CN117152579A

CN117152579A - System and computer-implemented method for a vehicle

Info

Publication number: CN117152579A
Application number: CN202211609528.XA
Authority: CN
Inventors: D·夏尔玛; V·E·B·里昂; S·A·维德贾亚; E·F·M·卡佩利耶
Original assignee: Motional AD LLC
Current assignee: Motional AD LLC
Priority date: 2022-06-01
Filing date: 2022-12-14
Publication date: 2023-12-01

Abstract

The invention relates to a system for a vehicle and a computer-implemented method. Methods, systems, and computer program products are provided for generating an output graph that indicates a likelihood that each element of an image corresponds to a particular road element (such as a lane separation line, a road separation line, and a road boundary). An example method may include: a machine learning architecture is applied to the image, wherein the architecture includes a convolutional neural network and a sub-network that captures global context from a feature map generated by the convolutional neural network.

Description

System and computer-implemented method for a vehicle

Technical Field

The present invention relates to systems and computer-implemented methods for vehicles, and in particular to automated lane marking extraction and classification from lidar scanning.

Disclosure of Invention

A computer-implemented method implemented by at least one processor, the computer-implemented method comprising: receiving, with the at least one processor, an image representing lidar scanned data from a vehicle environment; convolving, with the at least one processor, the image to generate a first feature map; converting, with the at least one processor, the first feature map into a first feature vector; applying, with the at least one processor, one or more additional convolutions to the first feature map to generate a second feature map; converting, with the at least one processor, the second feature map into a second feature vector; passing, with the at least one processor, an input representing at least the first feature vector and the second feature vector through a neural network to generate a scene feature vector; and generating, with the at least one processor, an output graph based at least on the scene feature vector and the second feature graph, wherein the output graph includes a plurality of pixels and indicates a likelihood that each pixel corresponds to a road element.

A system for a vehicle, comprising: one or more non-transitory data stores comprising computer-executable instructions; and one or more hardware processors configured to execute the computer-executable instructions to: receiving an image representing data from a lidar scan of a vehicle environment; convolving the image to generate a first feature map; converting the first feature map into a first feature vector; applying one or more additional convolutions to the first feature map to generate a second feature map; converting the second feature map into a second feature vector; passing an input representing at least the first feature vector and the second feature vector through a neural network to produce a scene feature vector; and generating an output graph based at least on the scene feature vector and the second feature graph, wherein the output graph includes a plurality of pixels and indicates a likelihood that each pixel corresponds to a road element.

A computer-implemented method, comprising: receiving, with at least one processor, data corresponding to an image of a lidar scan representing a vehicle environment; determining, with the at least one processor and using a machine learning model, a road element classification for a plurality of pixels of the image, wherein the classification for a particular pixel of the plurality of pixels indicates a traffic direction of a traffic lane associated with the particular pixel of the image; and generating, with the at least one processor, a plurality of multi-segment lines based on the classification for the plurality of pixels, wherein at least two of the plurality of multi-segment lines indicate a boundary between at least two traffic lanes for traffic traveling in different directions.

Drawings

FIG. 1 is an example environment in which a vehicle including one or more components of an autonomous system may be implemented;

FIG. 2 is a diagram of one or more systems of a vehicle including an autonomous system;

FIG. 3 is a diagram of components of one or more devices and/or one or more systems of FIGS. 1 and 2;

FIG. 4A is a diagram of certain components of an autonomous system;

FIG. 4B is a diagram of an implementation of a neural network;

fig. 4C and 4D are diagrams illustrating an example operation of the CNN;

FIG. 5 is a block diagram illustrating the operation of the lane classification system to accurately detect road elements from laser radar (lidar) data;

FIG. 6 is a visualization of an example image processed at various points during the operation of FIG. 5;

FIG. 7 is an exemplary machine learning architecture that may be implemented by the lane classification system of FIG. 5 to accurately detect road elements from lidar data;

FIG. 8 is an exemplary routine for classifying road elements within lidar scan data; and

FIG. 9 is an illustrative routine for image segmentation based on a convolutional neural network and a global context sub-network, which may be used to conduct classification as discussed with respect to FIG. 8.

Detailed Description

In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent, however, that the embodiments described in this disclosure may be practiced without these specific details. In some instances, well-known structures and devices are shown in block diagram form in order to avoid unnecessarily obscuring aspects of the present disclosure.

In the drawings, for ease of description, specific arrangements or sequences of illustrative elements (such as those representing systems, devices, modules, blocks of instructions, and/or data elements, etc.) are illustrated. However, those of skill in the art will understand that a specific order or arrangement of elements illustrated in the drawings is not intended to require a specific order or sequence of processes, or separation of processes, unless explicitly described. Furthermore, the inclusion of a schematic element in a figure is not intended to mean that such element is required in all embodiments nor that the feature represented by such element is not included in or combined with other elements in some embodiments unless explicitly described.

Furthermore, in the drawings, connecting elements (such as solid or dashed lines or arrows, etc.) are used to illustrate a connection, relationship or association between or among two or more other schematic elements, the absence of any such connecting element is not intended to mean that no connection, relationship or association exists. In other words, some connections, relationships, or associations between elements are not illustrated in the drawings so as not to obscure the present disclosure. Further, for ease of illustration, a single connection element may be used to represent multiple connections, relationships, or associations between elements. For example, if a connection element represents a communication of signals, data, or instructions (e.g., "software instructions"), those skilled in the art will understand that such element may represent one or more signal paths (e.g., buses) that may be required to effect the communication.

Although the terms "first," "second," and/or "third," etc. may be used to describe various elements, these elements should not be limited by these terms. The terms "first," second, "and/or third" are used merely to distinguish one element from another element. For example, a first contact may be referred to as a second contact, and similarly, a second contact may be referred to as a first contact, without departing from the scope of the described embodiments. Both the first contact and the second contact are contacts, but they are not the same contacts.

The terminology used in the description of the various embodiments described herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used in the specification of the various embodiments described and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, and may be used interchangeably with "one or more than one" or "at least one," unless the context clearly indicates otherwise. It will also be understood that the term "and/or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items. It will be further understood that the terms "comprises," "comprising," "includes," "including" and/or "having," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

As used herein, the terms "communication" and "communicating" refer to at least one of the receipt, transmission, and/or provision of information (or information represented by, for example, data, signals, messages, instructions, and/or commands, etc.). For one unit (e.g., a device, system, component of a device or system, and/or a combination thereof, etc.) to communicate with another unit, this means that the one unit is capable of directly or indirectly receiving information from and/or sending (e.g., transmitting) information to the other unit. This may refer to a direct or indirect connection that is wired and/or wireless in nature. In addition, two units may communicate with each other even though the transmitted information may be modified, processed, relayed and/or routed between the first unit and the second unit. For example, a first unit may communicate with a second unit even if the first unit passively receives information and does not actively transmit information to the second unit. As another example, if at least one intervening unit (e.g., a third unit located between the first unit and the second unit) processes information received from the first unit and transmits the processed information to the second unit, the first unit may communicate with the second unit. In some embodiments, a message may refer to a network packet (e.g., a data packet, etc.) that includes data.

As used herein, the term "if" is optionally interpreted to mean "when …", "at …", "in response to being determined to" and/or "in response to being detected", etc., depending on the context. Similarly, the phrase "if determined" or "if [ a stated condition or event ] is detected" is optionally interpreted to mean "upon determination …", "in response to determination" or "upon detection of [ a stated condition or event ]" and/or "in response to detection of [ a stated condition or event ]" or the like, depending on the context. Furthermore, as used herein, the terms "having," "having," or "owning," and the like, are intended to be open-ended terms. Furthermore, unless explicitly stated otherwise, the phrase "based on" is intended to mean "based, at least in part, on".

Reference will now be made in detail to embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the various embodiments described. It will be apparent, however, to one of ordinary skill in the art that the various embodiments described may be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.

General overview

Aspects of the present invention relate generally to generating highly accurate roadway maps representing road elements, such as lane divider lines (lane divider), road divider lines (road divider), and road boundaries, from lidar scans. More particularly, the present invention provides a machine learning model that can facilitate the generation of annotations to imaging data, such as road feature annotations to pseudo-images (pseudo-images) generated from lidar scans, and the like, using both Convolutional Neural Networks (CNNs) and scene feature vectors generated based on intermediate representations generated by the CNNs. As disclosed herein, the machine learning model provided herein may enable highly accurate annotations to be generated based on various features within portions of imaging data and contextual information provided by the data as a whole. In one example, embodiments of the present invention may be used to generate highly accurate road maps for guiding the operation of autonomous vehicles.

As will be appreciated by those skilled in the art, the safe and efficient operation of Autonomous Vehicles (AV) may generally depend to a large extent on accurate knowledge of the surrounding area. One mechanism for providing this knowledge is to provide a Bird's Eye View (BEV) map of the AV's operational area to the AV. For example, the AV may dynamically generate a map using the sensor data, or may obtain a map generated in advance. Objective accuracy may be very important, especially for pre-generated maps. That is, it may be critical for an AV to accurately understand its surroundings, and it may be critical for this understanding to reflect an objective true phase that is unchanged between AV (e.g., between an AV that created a map and an AV that is operating with the map).

Although various data capture techniques exist, many do not independently provide data of sufficient accuracy for map generation (particularly "off-line" map generation), where pre-generated maps are later provided to an operating vehicle for use as ground truth. For example, 2D imaging techniques such as conventional cameras may provide data that varies significantly based on location, viewpoint, environmental conditions, and camera parameters, among others. Thus, it may be difficult to generate a highly accurate map based on 2D imaging alone. Lidar, in contrast, has become an important technology for providing high accuracy understanding of the environment. In contrast to 2D imaging techniques, lidar additionally provides depth sensing, enabling 3D imaging of the region. These 3D images may provide accurate objective sensing in combination with accurate location information such as that provided by a Global Positioning System (GPS) or other location information. However, lidar may have some drawbacks relative to conventional 2D imaging such as cameras. For example, lidar may have difficulty distinguishing colors in sensed objects. As a result, it may be difficult to programmatically distinguish color-based identifications (such as lane separation lines, road separation lines, and road boundaries, etc.) from individual lidar data. While combining lidar data with other forms of data (such as camera data, etc.) may solve this problem, combining lidar data with additional modalities of data may not be a simple task in terms of engineering and operational complexity. For example, processing such a combination may require significantly more computational resources than processing lidar data alone. Accordingly, it may be beneficial to provide a system that facilitates high accuracy learning of features such as road identification from only lidar scans.

Embodiments of the present invention provide for this high accuracy learning of features such as road identification from lidar scanning (possibly alone or independent of data from other sensor modalities). Specifically, disclosed herein is a machine learning architecture that includes a CNN to process 2D image data (such as pseudo images generated from one or more lidar scans of an area, etc.) and identify features within the image data (such as road identifications, etc.). The road markings may include, for example, lane-dividing lines that divide lanes of the same direction on the roadway, lane-dividing line markings that divide lanes of different directions on the roadway, road boundaries that mark edges of the roadway, stop lines that indicate where the vehicle will stop, pedestrian markings that indicate a pedestrian path (e.g., a crosswalk), bike lane markings, and V-shaped markings. More particularly, the machine learning architecture disclosed herein may include a sub-network that generates scene feature vectors that capture context information of an entire environment from one or more intermediate representations generated by a CNN. As disclosed herein, the use of such sub-networks may enable more accurate differentiation between features that appear similar within the input data. For example, in the case of lidar data, lane and road separation lines (considering that lidar scans typically do not accurately capture color) may be similarly represented. Thus, existing machine learning models that typically focus specifically on the detection of specific local features may have difficulty distinguishing lane separation lines from road separation lines. However, by utilizing a sub-network providing scene feature vectors as disclosed herein, a model may be trained to more accurately distinguish these identifications based on context information reflecting the entire environment. Accordingly, the machine learning architecture disclosed herein may represent an improvement over the prior art for identifying features within imaging data, thereby enabling more accurate generation of maps that may be used in various contexts (such as AV, etc.).

Referring now to FIG. 1, an example environment 100 is illustrated in which a vehicle that includes an autonomous system and a vehicle that does not include an autonomous system operate in the example environment 100. As illustrated, environment 100 includes vehicles 102a-102n, objects 104a-104n, routes 106a-106n, areas 108, vehicle-to-infrastructure (V2I) devices 110, a network 112, a remote Autonomous Vehicle (AV) system 114, a queue management system 116, and a V2I system 118. The vehicles 102a-102n, the vehicle-to-infrastructure (V2I) device 110, the network 112, the Autonomous Vehicle (AV) system 114, the queue management system 116, and the V2I system 118 are interconnected via wired connections, wireless connections, or a combination of wired or wireless connections (e.g., establishing a connection for communication, etc.). In some embodiments, the objects 104a-104n are interconnected with at least one of the vehicles 102a-102n, the vehicle-to-infrastructure (V2I) device 110, the network 112, the Autonomous Vehicle (AV) system 114, the queue management system 116, and the V2I system 118 via a wired connection, a wireless connection, or a combination of wired or wireless connections.

The vehicles 102a-102n (individually referred to as vehicles 102 and collectively referred to as vehicles 102) include at least one device configured to transport cargo and/or personnel. In some embodiments, the vehicle 102 is configured to communicate with the V2I device 110, the remote AV system 114, the queue management system 116, and/or the V2I system 118 via the network 112. In some embodiments, the vehicle 102 comprises a car, bus, truck, train, or the like. In some embodiments, the vehicle 102 is the same as or similar to the vehicle 200 (see fig. 2) described herein. In some embodiments, vehicles 200 in the collection of vehicles 200 are associated with an autonomous queue manager. In some embodiments, the vehicles 102 travel along respective routes 106a-106n (individually referred to as routes 106 and collectively referred to as routes 106), as described herein. In some embodiments, one or more vehicles 102 include an autonomous system (e.g., the same or similar to autonomous system 202).

The objects 104a-104n (individually referred to as objects 104 and collectively referred to as objects 104) include, for example, at least one vehicle, at least one pedestrian, at least one rider, and/or at least one structure (e.g., building, sign, hydrant, etc.), and the like. Each object 104 is stationary (e.g., at a fixed location and for a period of time) or moves (e.g., has a velocity and is associated with at least one trajectory). In some embodiments, the object 104 is associated with a respective location in the region 108.

Routes 106a-106n (individually referred to as routes 106 and collectively referred to as routes 106) are each associated with (e.g., define) a series of actions (also referred to as tracks) that connect the states along which the AV can navigate. Each route 106 begins in an initial state (e.g., a state corresponding to a first space-time location and/or speed, etc.) and ends in a final target state (e.g., a state corresponding to a second space-time location different from the first space-time location) or target area (e.g., a subspace of acceptable states (e.g., end states)). In some embodiments, the first state includes one or more places where the one or more individuals are to pick up the AV, and the second state or zone includes one or more places where the one or more individuals pick up the AV are to be off. In some embodiments, the route 106 includes a plurality of acceptable state sequences (e.g., a plurality of spatiotemporal site sequences) associated with (e.g., defining) a plurality of trajectories. In an example, the route 106 includes only high-level actions or imprecise status places, such as a series of connecting roads indicating a change of direction at a roadway intersection, and the like. Additionally or alternatively, the route 106 may include more precise actions or states such as, for example, specific target lanes or precise locations within a lane region, and target speeds at these locations, etc. In an example, the route 106 includes a plurality of precise state sequences along at least one high-level action with a limited look-ahead view to an intermediate target, where a combination of successive iterations of the limited view state sequences cumulatively corresponds to a plurality of trajectories that collectively form a high-level route that terminates at a final target state or zone.

The area 108 includes a physical area (e.g., a geographic area) that the vehicle 102 may navigate. In an example, the region 108 includes at least one state (e.g., a country, a province, an individual state of a plurality of states included in a country, etc.), at least a portion of a state, at least one city, at least a portion of a city, etc. In some embodiments, the area 108 includes at least one named thoroughfare (referred to herein as a "road"), such as a highway, interstate, park, city street, or the like. Additionally or alternatively, in some examples, the area 108 includes at least one unnamed road, such as a roadway, a section of a parking lot, a section of an open space and/or undeveloped area, a mud path, and the like. In some embodiments, the roadway includes at least one lane (e.g., a portion of the roadway through which the vehicle 102 may traverse). In an example, the road includes at least one lane associated with (e.g., identified based on) the at least one lane identification.

A Vehicle-to-infrastructure (V2I) device 110 (sometimes referred to as a Vehicle-to-Everything (V2X) device) includes at least one device configured to communicate with the Vehicle 102 and/or the V2I system 118. In some embodiments, V2I device 110 is configured to communicate with vehicle 102, remote AV system 114, queue management system 116, and/or V2I system 118 via network 112. In some embodiments, V2I device 110 includes a Radio Frequency Identification (RFID) device, a sign, a camera (e.g., a two-dimensional (2D) and/or three-dimensional (3D) camera), a lane marker, a street light, a parking meter, and the like. In some embodiments, the V2I device 110 is configured to communicate directly with the vehicle 102. Additionally or alternatively, in some embodiments, the V2I device 110 is configured to communicate with the vehicle 102, the remote AV system 114, and/or the queue management system 116 via the V2I system 118. In some embodiments, V2I device 110 is configured to communicate with V2I system 118 via network 112.

Network 112 includes one or more wired and/or wireless networks. In an example, the network 112 includes a cellular network (e.g., a Long Term Evolution (LTE) network, a third generation (3G) network, a fourth generation (4G) network, a fifth generation (5G) network, a Code Division Multiple Access (CDMA) network, etc.), a Public Land Mobile Network (PLMN), a Local Area Network (LAN), a Wide Area Network (WAN), a Metropolitan Area Network (MAN), a telephone network (e.g., a Public Switched Telephone Network (PSTN)), a private network, an ad hoc network, an intranet, the internet, a fiber-optic based network, a cloud computing network, etc., and/or a combination of some or all of these networks, etc.

The remote AV system 114 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the network 112, the queue management system 116, and/or the V2I system 118 via the network 112. In an example, the remote AV system 114 includes a server, a group of servers, and/or other similar devices. In some embodiments, the remote AV system 114 is co-located with the queue management system 116. In some embodiments, the remote AV system 114 participates in the installation of some or all of the components of the vehicle (including autonomous systems, autonomous vehicle computing, and/or software implemented by autonomous vehicle computing, etc.). In some embodiments, the remote AV system 114 maintains (e.g., updates and/or replaces) these components and/or software over the life of the vehicle.

The queue management system 116 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the remote AV system 114, and/or the V2I system 118. In an example, the queue management system 116 includes a server, a server group, and/or other similar devices. In some embodiments, the queue management system 116 is associated with a carpool company (e.g., an organization for controlling operation of multiple vehicles (e.g., vehicles that include autonomous systems and/or vehicles that do not include autonomous systems), etc.).

In some embodiments, the V2I system 118 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the remote AV system 114, and/or the queue management system 116 via the network 112. In some examples, the V2I system 118 is configured to communicate with the V2I device 110 via a connection other than the network 112. In some embodiments, V2I system 118 includes a server, a server farm, and/or other similar devices. In some embodiments, the V2I system 118 is associated with a municipality or private institution (e.g., a private institution for maintaining the V2I device 110, etc.).

The number and arrangement of elements illustrated in fig. 1 are provided as examples. There may be additional elements, fewer elements, different elements, and/or differently arranged elements than those illustrated in fig. 1. Additionally or alternatively, at least one element of environment 100 may perform one or more functions described as being performed by at least one different element of fig. 1. Additionally or alternatively, at least one set of elements of environment 100 may perform one or more functions described as being performed by at least one different set of elements of environment 100.

Referring now to fig. 2, a vehicle 200 includes an autonomous system 202, a powertrain control system 204, a steering control system 206, and a braking system 208. In some embodiments, the vehicle 200 is the same as or similar to the vehicle 102 (see fig. 1). In some embodiments, vehicle 200 has autonomous capabilities (e.g., implements at least one function, feature, and/or means, etc., that enables vehicle 200 to operate partially or fully without human intervention, including, but not limited to, a fully autonomous vehicle (e.g., a vehicle that foregoes human intervention), and/or a highly autonomous vehicle (e.g., a vehicle that foregoes human intervention in some cases), etc. For a detailed description of fully autonomous vehicles and highly autonomous vehicles, reference may be made to SAE International Standard J3016, classification and definition of on-road automotive autopilot system related terms (SAE International's Standard J3016: taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems), which is incorporated by reference in its entirety. In some embodiments, the vehicle 200 is associated with an autonomous queue manager and/or a carpooling company.

The autonomous system 202 includes a sensor suite that includes one or more devices such as a camera 202a, liDAR sensor 202b, radar (radar) sensor 202c, and microphone 202 d. In some embodiments, autonomous system 202 may include more or fewer devices and/or different devices (e.g., ultrasonic sensors, inertial sensors, GPS receivers (discussed below), and/or odometry sensors for generating data associated with an indication of the distance that vehicle 200 has traveled, etc.). In some embodiments, the autonomous system 202 uses one or more devices included in the autonomous system 202 to generate data associated with the environment 100 described herein. The data generated by the one or more devices of the autonomous system 202 may be used by the one or more systems described herein to observe the environment (e.g., environment 100) in which the vehicle 200 is located. In some embodiments, autonomous system 202 includes a communication device 202e, an autonomous vehicle calculation 202f, and a safety controller 202g.

The camera 202a includes at least one device configured to communicate with the communication device 202e, the autonomous vehicle calculation 202f, and/or the safety controller 202g via a bus (e.g., the same or similar to the bus 302 of fig. 3). The camera 202a includes at least one camera (e.g., a digital camera using a light sensor such as a Charge Coupled Device (CCD), thermal camera, infrared (IR) camera, event camera, etc.) to capture images including physical objects (e.g., cars, buses, curbs, and/or people, etc.). In some embodiments, camera 202a generates camera data as output. In some examples, camera 202a generates camera data including image data associated with the image. In this example, the image data may specify at least one parameter corresponding to the image (e.g., image characteristics such as exposure, brightness, etc., and/or an image timestamp, etc.). In such examples, the image may be in a format (e.g., RAW, JPEG, and/or PNG, etc.). In some embodiments, the camera 202a includes a plurality of independent cameras configured (e.g., positioned) on the vehicle to capture images for stereoscopic (stereo vision) purposes. In some examples, camera 202a includes a plurality of cameras that generate and transmit image data to autonomous vehicle computing 202f and/or a queue management system (e.g., a queue management system that is the same as or similar to queue management system 116 of fig. 1). In such an example, the autonomous vehicle calculation 202f determines a depth to one or more objects in the field of view of at least two cameras of the plurality of cameras based on image data from the at least two cameras. In some embodiments, camera 202a is configured to capture images of objects within a distance (e.g., up to 100 meters and/or up to 1 kilometer, etc.) relative to camera 202 a. Thus, the camera 202a includes features such as sensors and lenses that are optimized for sensing objects at one or more distances relative to the camera 202 a.

In an embodiment, camera 202a includes at least one camera configured to capture one or more images associated with one or more traffic lights, street signs, and/or other physical objects that provide visual navigation information. In some embodiments, the camera 202a generates traffic light data (TLD data) associated with one or more images. In some examples, the camera 202a generates TLD data associated with one or more images including formats (e.g., RAW, JPEG, and/or PNG, etc.). In some embodiments, the camera 202a that generates TLD data differs from other systems described herein that include cameras in that: the camera 202a may include one or more cameras having a wide field of view (e.g., wide angle lens, fisheye lens, and/or lens having a viewing angle of about 120 degrees or greater, etc.) to generate images related to as many physical objects as possible.

Laser detection and ranging (LiDAR) sensor 202b includes at least one device configured to communicate with communication device 202e, autonomous vehicle computation 202f, and/or security controller 202g via a bus (e.g., the same or similar bus as bus 302 of fig. 3). LiDAR sensor 202b includes a system configured to emit light from a light emitter (e.g., a laser emitter). Light emitted by the LiDAR sensor 202b includes light outside the visible spectrum (e.g., infrared light, etc.). In some embodiments, during operation, light emitted by the LiDAR sensor 202b encounters a physical object (e.g., a vehicle) and is reflected back to the LiDAR sensor 202b. In some embodiments, the light emitted by LiDAR sensor 202b does not penetrate the physical object that the light encounters. LiDAR sensor 202b also includes at least one light detector that detects light emitted from the light emitter after the light encounters a physical object. In some embodiments, at least one data processing system associated with the LiDAR sensor 202b generates an image (e.g., a point cloud and/or a combined point cloud, etc.) representative of objects included in the field of view of the LiDAR sensor 202b. In some examples, at least one data processing system associated with the LiDAR sensor 202b generates images representing boundaries of the physical object and/or surfaces (e.g., topology of surfaces) of the physical object, etc. In such an example, the image is used to determine the boundary of a physical object in the field of view of the LiDAR sensor 202b.

The radio detection and ranging (radar) sensor 202c includes at least one device configured to communicate with the communication device 202e, the autonomous vehicle calculation 202f, and/or the safety controller 202g via a bus (e.g., the same or similar bus as the bus 302 of fig. 3). The radar sensor 202c includes a system configured to emit (pulsed or continuous) radio waves. The radio waves emitted by the radar sensor 202c include radio waves within a predetermined frequency spectrum. In some embodiments, during operation, radio waves emitted by the radar sensor 202c encounter a physical object and are reflected back to the radar sensor 202c. In some embodiments, the radio waves emitted by the radar sensor 202c are not reflected by some objects. In some embodiments, at least one data processing system associated with radar sensor 202c generates signals representative of objects included in the field of view of radar sensor 202c. For example, at least one data processing system associated with radar sensor 202c generates images representing boundaries of physical objects and/or surfaces (e.g., topology of surfaces) of physical objects, etc. In some examples, the image is used to determine boundaries of physical objects in the field of view of radar sensor 202c.

Microphone 202d includes at least one device configured to communicate with communication device 202e, autonomous vehicle computing 202f, and/or security controller 202g via a bus (e.g., the same or similar bus as bus 302 of fig. 3). Microphone 202d includes one or more microphones (e.g., array microphone and/or external microphone, etc.) that capture an audio signal and generate data associated with (e.g., representative of) the audio signal. In some examples, microphone 202d includes transducer means and/or the like. In some embodiments, one or more systems described herein may receive data generated by microphone 202d and determine a position (e.g., distance, etc.) of an object relative to vehicle 200 based on an audio signal associated with the data.

The communication device 202e includes at least one device configured to communicate with a camera 202a, a LiDAR sensor 202b, a radar sensor 202c, a microphone 202d, an autonomous vehicle calculation 202f, a security controller 202g, and/or a drive-by-wire (DBW) system 202 h. For example, communication device 202e may include the same or similar devices as communication interface 314 of fig. 3. In some embodiments, the communication device 202e comprises a vehicle-to-vehicle (V2V) communication device (e.g., a device for enabling wireless communication of data between vehicles).

The autonomous vehicle calculation 202f includes at least one device configured to communicate with the camera 202a, the LiDAR sensor 202b, the radar sensor 202c, the microphone 202d, the communication device 202e, the security controller 202g, and/or the DBW system 202 h. In some examples, the autonomous vehicle computing 202f includes devices such as client devices, mobile devices (e.g., cellular phones and/or tablet computers, etc.), and/or servers (e.g., computing devices including one or more central processing units and/or graphics processing units, etc.), among others. In some embodiments, the autonomous vehicle calculation 202f is the same as or similar to the autonomous vehicle calculation 400 described herein. Additionally or alternatively, in some embodiments, the autonomous vehicle computing 202f is configured to communicate with an autonomous vehicle system (e.g., an autonomous vehicle system that is the same as or similar to the remote AV system 114 of fig. 1), a queue management system (e.g., a queue management system that is the same as or similar to the queue management system 116 of fig. 1), a V2I device (e.g., a V2I device that is the same as or similar to the V2I device 110 of fig. 1), and/or a V2I system (e.g., a V2I system that is the same as or similar to the V2I system 118 of fig. 1).

The safety controller 202g includes at least one device configured to communicate with the camera 202a, the LiDAR sensor 202b, the radar sensor 202c, the microphone 202d, the communication device 202e, the autonomous vehicle calculation 202f, and/or the DBW system 202 h. In some examples, the safety controller 202g includes one or more controllers (electrical and/or electromechanical controllers, etc.) configured to generate and/or transmit control signals to operate one or more devices of the vehicle 200 (e.g., the powertrain control system 204, the steering control system 206, and/or the braking system 208, etc.). In some embodiments, the safety controller 202g is configured to generate control signals that override (e.g., override) control signals generated and/or transmitted by the autonomous vehicle calculation 202 f.

The DBW system 202h includes at least one device configured to communicate with the communication device 202e and/or the autonomous vehicle calculation 202 f. In some examples, the DBW system 202h includes one or more controllers (e.g., electrical and/or electromechanical controllers, etc.) configured to generate and/or transmit control signals to operate one or more devices of the vehicle 200 (e.g., the powertrain control system 204, the steering control system 206, and/or the braking system 208, etc.). Additionally or alternatively, one or more controllers of the DBW system 202h are configured to generate and/or transmit control signals to operate at least one different device of the vehicle 200 (e.g., turn signal lights, headlights, door locks, and/or windshield wipers, etc.).

The powertrain control system 204 includes at least one device configured to communicate with the DBW system 202 h. In some examples, the powertrain control system 204 includes at least one controller and/or actuator, etc. In some embodiments, the powertrain control system 204 receives control signals from the DBW system 202h, and the powertrain control system 204 causes the vehicle 200 to begin moving forward, stop moving forward, begin moving backward, stop moving backward, accelerate in a direction, decelerate in a direction, make a left turn, make a right turn, and/or the like. In an example, the powertrain control system 204 increases, maintains the same, or decreases the energy (e.g., fuel and/or electricity, etc.) provided to the motor of the vehicle, thereby rotating or not rotating at least one wheel of the vehicle 200.

The steering control system 206 includes at least one device configured to rotate one or more wheels of the vehicle 200. In some examples, the steering control system 206 includes at least one controller and/or actuator, etc. In some embodiments, steering control system 206 rotates the two front wheels and/or the two rear wheels of vehicle 200 to the left or right to turn vehicle 200 to the left or right.

The braking system 208 includes at least one device configured to actuate one or more brakes to slow and/or hold the vehicle 200 stationary. In some examples, the braking system 208 includes at least one controller and/or actuator configured to cause one or more calipers associated with one or more wheels of the vehicle 200 to close on a respective rotor of the vehicle 200. Additionally or alternatively, in some examples, the braking system 208 includes an Automatic Emergency Braking (AEB) system and/or a regenerative braking system, or the like.

In some embodiments, the vehicle 200 includes at least one platform sensor (not explicitly illustrated) for measuring or inferring a property of the state or condition of the vehicle 200. In some examples, the vehicle 200 includes platform sensors such as a Global Positioning System (GPS) receiver, an Inertial Measurement Unit (IMU), a wheel speed sensor, a wheel brake pressure sensor, a wheel torque sensor, an engine torque sensor, and/or a steering angle sensor, among others.

Referring now to fig. 3, a schematic diagram of an apparatus 300 is illustrated. As illustrated, the apparatus 300 includes a processor 304, a memory 306, a storage component 308, an input interface 310, an output interface 312, a communication interface 314, and a bus 302. In some embodiments, the apparatus 300 corresponds to: at least one device of the vehicle 102 (e.g., at least one device of a system of the vehicle 102); and/or one or more devices of network 112 (e.g., one or more devices of a system of network 112). In some embodiments, one or more devices of the vehicle 102 (e.g., one or more devices of the system of the vehicle 102), and/or one or more devices of the network 112 (e.g., one or more devices of the system of the network 112) include at least one device 300 and/or at least one component of the device 300. As shown in fig. 3, the apparatus 300 includes a bus 302, a processor 304, a memory 306, a storage component 308, an input interface 310, an output interface 312, and a communication interface 314.

Bus 302 includes components that permit communication between the components of device 300. In some cases, processor 304 includes a processor (e.g., a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and/or an Acceleration Processing Unit (APU), etc.), a microphone, a Digital Signal Processor (DSP), and/or any processing component that may be programmed to perform at least one function (e.g., a Field Programmable Gate Array (FPGA), and/or an Application Specific Integrated Circuit (ASIC), etc.). Memory 306 includes Random Access Memory (RAM), read Only Memory (ROM), and/or another type of dynamic and/or static storage device (e.g., flash memory, magnetic and/or optical memory, etc.) that stores data and/or instructions for use by processor 304.

The storage component 308 stores data and/or software related to operation and use of the apparatus 300. In some examples, storage component 308 includes a hard disk (e.g., magnetic disk, optical disk, magneto-optical disk, and/or solid state disk, etc.), a Compact Disk (CD), a Digital Versatile Disk (DVD), a floppy disk, a magnetic cassette tape, a magnetic tape, a CD-ROM, RAM, PROM, EPROM, FLASH-EPROM, NV-RAM, and/or another type of computer-readable medium, and a corresponding drive.

Input interface 310 includes components that permit device 300 to receive information, such as via user input (e.g., a touch screen display, keyboard, keypad, mouse, buttons, switches, microphone, and/or camera, etc.). Additionally or alternatively, in some embodiments, the input interface 310 includes sensors (e.g., global Positioning System (GPS) receivers, accelerometers, gyroscopes, and/or actuators, etc.) for sensing information. Output interface 312 includes components (e.g., a display, a speaker, and/or one or more Light Emitting Diodes (LEDs), etc.) for providing output information from device 300.

In some embodiments, the communication interface 314 includes transceiver-like components (e.g., a transceiver and/or separate receivers and transmitters, etc.) that permit the device 300 to communicate with other devices via a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection. In some examples, the communication interface 314 permits the device 300 to receive information from and/or provide information to another device. In some of the examples of the present invention, communication interface 314 includes an ethernet interface, an optical interface, a coaxial interface an infrared interface, a Radio Frequency (RF) interface, a Universal Serial Bus (USB) interface, An interface and/or a cellular network interface, etc.

In some embodiments, the apparatus 300 performs one or more of the processes described herein. The apparatus 300 performs these processes based on the processor 304 executing software instructions stored by a computer readable medium, such as the memory 306 and/or the storage component 308. A computer-readable medium (e.g., a non-transitory computer-readable medium) is defined herein as a non-transitory memory device. Non-transitory memory devices include storage space located within a single physical storage device or distributed across multiple physical storage devices.

In some embodiments, the software instructions are read into memory 306 and/or storage component 308 from another computer-readable medium or from another device via communication interface 314. The software instructions stored in memory 306 and/or storage component 308, when executed, cause processor 304 to perform one or more of the processes described herein. Additionally or alternatively, hardwired circuitry is used in place of or in combination with software instructions to perform one or more processes described herein. Thus, unless explicitly stated otherwise, the embodiments described herein are not limited to any specific combination of hardware circuitry and software.

Memory 306 and/or storage component 308 includes a data store or at least one data structure (e.g., database, etc.). The apparatus 300 is capable of receiving information from, storing information in, communicating information to, or searching information stored in a data store or at least one data structure in the memory 306 or storage component 308. In some examples, the information includes network data, input data, output data, or any combination thereof.

In some embodiments, apparatus 300 is configured to execute software instructions stored in memory 306 and/or a memory of another apparatus (e.g., another apparatus that is the same as or similar to apparatus 300). As used herein, the term "module" refers to at least one instruction stored in memory 306 and/or a memory of another device that, when executed by processor 304 and/or a processor of another device (e.g., another device that is the same as or similar to device 300), causes device 300 (e.g., at least one component of device 300) to perform one or more processes described herein. In some embodiments, the modules are implemented in software, firmware, hardware, and/or the like.

The number and arrangement of components illustrated in fig. 3 are provided as examples. In some embodiments, apparatus 300 may include additional components, fewer components, different components, or differently arranged components than those illustrated in fig. 3. Additionally or alternatively, a set of components (e.g., one or more components) of the apparatus 300 may perform one or more functions described as being performed by another component or another set of components of the apparatus 300.

Referring now to fig. 4A, an example block diagram of an autonomous vehicle computation 400 (sometimes referred to as an "AV stack") is illustrated. As illustrated, autonomous vehicle computation 400 includes a perception system 402 (sometimes referred to as a perception module), a planning system 404 (sometimes referred to as a planning module), a positioning system 406 (sometimes referred to as a positioning module), a control system 408 (sometimes referred to as a control module), and a database 410. In some embodiments, the perception system 402, the planning system 404, the positioning system 406, the control system 408, and the database 410 are included in and/or implemented in an automated navigation system of the vehicle (e.g., the autonomous vehicle calculation 202f of the vehicle 200). Additionally or alternatively, in some embodiments, the perception system 402, the planning system 404, the positioning system 406, the control system 408, and the database 410 are included in one or more independent systems (e.g., one or more systems identical or similar to the autonomous vehicle calculation 400, etc.). In some examples, the perception system 402, the planning system 404, the positioning system 406, the control system 408, and the database 41 are included in one or more independent systems located in the vehicle and/or at least one remote system as described herein. In some embodiments, any and/or all of the systems included in autonomous vehicle computing 400 are implemented in software (e.g., software instructions stored in memory), computer hardware (e.g., by microprocessors, microcontrollers, application Specific Integrated Circuits (ASICs), and/or Field Programmable Gate Arrays (FPGAs), etc.), or a combination of computer software and computer hardware. It will also be appreciated that in some embodiments, the autonomous vehicle computing 400 is configured to communicate with a remote system (e.g., an autonomous vehicle system that is the same as or similar to the remote AV system 114, a queue management system 116 that is the same as or similar to the queue management system 116, and/or a V2I system that is the same as or similar to the V2I system 118, etc.).

In some embodiments, the perception system 402 receives data associated with at least one physical object in the environment (e.g., data used by the perception system 402 to detect the at least one physical object) and classifies the at least one physical object. In some examples, perception system 402 receives image data captured by at least one camera (e.g., camera 202 a) that is associated with (e.g., represents) one or more physical objects within a field of view of the at least one camera. In such examples, the perception system 402 classifies at least one physical object based on one or more groupings of physical objects (e.g., bicycles, vehicles, traffic signs, and/or pedestrians, etc.). In some embodiments, based on the classification of the physical object by the perception system 402, the perception system 402 transmits data associated with the classification of the physical object to the planning system 404.

In some embodiments, planning system 404 receives data associated with a destination and generates data associated with at least one route (e.g., route 106) along which a vehicle (e.g., vehicle 102) may travel toward the destination. In some embodiments, the planning system 404 receives data (e.g., the data associated with the classification of the physical object described above) from the perception system 402 periodically or continuously, and the planning system 404 updates at least one trajectory or generates at least one different trajectory based on the data generated by the perception system 402. In some embodiments, planning system 404 receives data associated with an updated position of a vehicle (e.g., vehicle 102) from positioning system 406, and planning system 404 updates at least one track or generates at least one different track based on the data generated by positioning system 406.

In some embodiments, the positioning system 406 receives data associated with (e.g., representative of) a location of a vehicle (e.g., the vehicle 102) in an area. In some examples, the positioning system 406 receives LiDAR data associated with at least one point cloud generated by at least one LiDAR sensor (e.g., liDAR sensor 202 b). In some examples, the positioning system 406 receives data associated with at least one point cloud from a plurality of LiDAR sensors, and the positioning system 406 generates a combined point cloud based on each point cloud. In these examples, the positioning system 406 compares the at least one point cloud or combined point cloud to a two-dimensional (2D) and/or three-dimensional (3D) map of the area stored in the database 410. The location system 406 then determines the location of the vehicle in the area based on the location system 406 comparing the at least one point cloud or combined point cloud to the map. In some embodiments, the map includes a combined point cloud for the region generated prior to navigation of the vehicle. In some embodiments, the map includes, but is not limited to, a high-precision map of roadway geometry, a map describing road network connection properties, a map describing roadway physical properties (such as traffic rate, traffic flow, number of vehicles and bicycle traffic lanes, lane width, type and location of lane traffic direction or lane markings, or combinations thereof, etc.), and a map describing spatial locations of roadway features (such as crosswalks, traffic signs or various types of other travel signals, etc.). In some embodiments, the map is generated in real-time based on data received by the perception system.

In another example, the positioning system 406 receives Global Navigation Satellite System (GNSS) data generated by a Global Positioning System (GPS) receiver. In some examples, positioning system 406 receives GNSS data associated with a location of a vehicle in an area, and positioning system 406 determines a latitude and longitude of the vehicle in the area. In such examples, the positioning system 406 determines the location of the vehicle in the area based on the latitude and longitude of the vehicle. In some embodiments, the positioning system 406 generates data associated with the position of the vehicle. In some examples, based on the positioning system 406 determining the location of the vehicle, the positioning system 406 generates data associated with the location of the vehicle. In such examples, the data associated with the location of the vehicle includes data associated with one or more semantic properties corresponding to the location of the vehicle.

In some embodiments, control system 408 receives data associated with at least one trajectory from planning system 404, and control system 408 controls the operation of the vehicle. In some examples, the control system 408 receives data associated with at least one trajectory from the planning system 404, and the control system 408 controls operation of the vehicle by generating and transmitting control signals to operate a powertrain control system (e.g., the DBW system 202h and/or the powertrain control system 204, etc.), a steering control system (e.g., the steering control system 206), and/or a braking system (e.g., the braking system 208). In an example, where the trajectory includes a left turn, the control system 408 transmits a control signal to cause the steering control system 206 to adjust the steering angle of the vehicle 200, thereby causing the vehicle 200 to turn left. Additionally or alternatively, the control system 408 generates and transmits control signals to cause other devices of the vehicle 200 (e.g., headlights, turn signal lights, door locks, and/or windshield wipers, etc.) to change state.

In some embodiments, the perception system 402, the planning system 404, the localization system 406, and/or the control system 408 implement at least one machine learning model (e.g., at least one multi-layer perceptron (MLP), at least one Convolutional Neural Network (CNN), at least one Recurrent Neural Network (RNN), at least one automatic encoder and/or at least one transformer, etc.). In some examples, the perception system 402, the planning system 404, the positioning system 406, and/or the control system 408 implement at least one machine learning model alone or in combination with one or more of the above systems. In some examples, the perception system 402, the planning system 404, the positioning system 406, and/or the control system 408 implement at least one machine learning model as part of a pipeline (e.g., a pipeline for identifying one or more objects located in an environment, etc.). Examples of implementations of the machine learning model are included below with respect to fig. 4B-4D.

Database 410 stores data transmitted to, received from, and/or updated by sensing system 402, planning system 404, positioning system 406, and/or control system 408. In some examples, database 410 includes a storage component (e.g., the same or similar to storage component 308 of fig. 3) for storing data and/or software related to operations and using at least one system of autonomous vehicle computing 400. In some embodiments, database 410 stores data associated with 2D and/or 3D maps of at least one region. In some examples, database 410 stores data associated with 2D and/or 3D maps of a portion of a city, portions of multiple cities, counties, states, and/or countries (states) (e.g., countries), etc. In such examples, a vehicle (e.g., the same or similar vehicle as vehicle 102 and/or vehicle 200) may drive along one or more drivable regions (e.g., single lane roads, multi-lane roads, highways, remote roads, and/or off-road roads, etc.) and cause at least one LiDAR sensor (e.g., the same or similar LiDAR sensor as LiDAR sensor 202 b) to generate data associated with an image representative of an object included in a field of view of the at least one LiDAR sensor.

In some embodiments, database 410 may be implemented across multiple devices. In some examples, database 410 is included in a vehicle (e.g., the same or similar to vehicle 102 and/or vehicle 200), an autonomous vehicle system (e.g., the same or similar to remote AV system 114), a queue management system (e.g., the same or similar to queue management system 116 of fig. 1), and/or a V2I system (e.g., the same or similar to V2I system 118 of fig. 1), etc.

Referring now to FIG. 4B, a diagram of an implementation of a machine learning model is illustrated. More specifically, a diagram illustrating an implementation of Convolutional Neural Network (CNN) 420. For purposes of illustration, the following description of CNN 420 will be with respect to the implementation of CNN 420 by sensing system 402. However, it will be appreciated that in some examples, CNN 420 (e.g., one or more components of CNN 420) is implemented by other systems (such as planning system 404, positioning system 406, and/or control system 408, etc.) other than sensing system 402 or in addition to sensing system 402. Although CNN 420 includes certain features as described herein, these features are provided for illustrative purposes and are not intended to limit the present disclosure.

CNN420 includes a plurality of convolutional layers including a first convolutional layer 422, a second convolutional layer 424, and a convolutional layer 426. In some embodiments, CNN420 includes a sub-sampling layer 428 (sometimes referred to as a pooling layer). In some embodiments, the sub-sampling layer 428 and/or other sub-sampling layers have dimensions that are smaller than the dimensions of the upstream system (i.e., the amount of nodes). By means of the sub-sampling layer 428 having a dimension smaller than that of the upstream layer, the CNN420 merges the amount of data associated with the initial input and/or output of the upstream layer, thereby reducing the amount of computation required by the CNN420 to perform the downstream convolution operation. Additionally or alternatively, CNN420 incorporates the amount of data associated with the initial input by way of sub-sampling layer 428 being associated with (e.g., configured to perform) at least one sub-sampling function (as described below with respect to fig. 4C and 4D).

Based on the perception system 402 providing respective inputs and/or outputs associated with each of the first convolution layer 422, the second convolution layer 424, and the convolution layer 426 to generate respective outputs, the perception system 402 performs convolution operations. In some examples, the perception system 402 implements the CNN420 based on the perception system 402 providing data as input to a first convolution layer 422, a second convolution layer 424, and a convolution layer 426. In such examples, based on the perception system 402 receiving data from one or more different systems (e.g., one or more systems of a vehicle that is the same or similar to the vehicle 102, a remote AV system that is the same or similar to the remote AV system 114, a queue management system that is the same or similar to the queue management system 116, and/or a V2I system that is the same or similar to the V2I system 118, etc.), the perception system 402 provides data as input to the first convolution layer 422, the second convolution layer 424, and the convolution layer 426. The following detailed description of the convolution operation is included with respect to fig. 4C.

In some embodiments, the perception system 402 provides data associated with an input (referred to as an initial input) to a first convolution layer 422, and the perception system 402 generates data associated with an output using the first convolution layer 422. In some embodiments, the perception system 402 provides as input the output generated by the convolutional layers to the different convolutional layers. For example, the perception system 402 provides the output of the first convolution layer 422 as an input to the sub-sampling layer 428, the second convolution layer 424, and/or the convolution layer 426. In such examples, the first convolution layer 422 is referred to as an upstream layer and the sub-sampling layer 428, the second convolution layer 424, and/or the convolution layer 426 are referred to as downstream layers. Similarly, in some embodiments, the perception system 402 provides the output of the sub-sampling layer 428 to the second convolution layer 424 and/or the convolution layer 426, and in this example, the sub-sampling layer 428 will be referred to as an upstream layer and the second convolution layer 424 and/or the convolution layer 426 will be referred to as a downstream layer.

In some embodiments, the perception system 402 processes data associated with the input provided to the CNN 420 before the perception system 402 provides the input to the CNN 420. For example, based on the sensor data (e.g., image data, liDAR data, radar data, etc.) being normalized by the perception system 402, the perception system 402 processes data associated with the input provided to the CNN 420.

In some embodiments, CNN 420 generates an output based on the perceptual system 402 performing convolution operations associated with the respective convolution layers. In some examples, CNN 420 generates an output based on the perception system 402 performing convolution operations associated with the various convolution layers and the initial input. In some embodiments, the perception system 402 generates an output and provides the output to the fully connected layer 430. In some examples, the perception system 402 provides the output of the convolutional layer 426 to the fully-connected layer 430, where the fully-connected layer 430 includes data associated with a plurality of characteristic values referred to as F1, F2. In this example, the output of convolution layer 426 includes data associated with a plurality of output characteristic values representing predictions.

In some embodiments, based on the perception system 402 identifying the feature value associated with the highest likelihood as the correct prediction of the plurality of predictions, the perception system 402 identifies the prediction from the plurality of predictions. For example, where fully connected layer 430 includes eigenvalues F1, F2,..fn, and F1 is the largest eigenvalue, perception system 402 identifies the prediction associated with F1 as the correct prediction of the plurality of predictions. In some embodiments, the perception system 402 trains the CNN 420 to generate predictions. In some examples, based on perception system 402 providing training data associated with the predictions to CNN 420, perception system 402 trains CNN 420 to generate the predictions.

Referring now to fig. 4C and 4D, diagrams illustrating example operations of CNN440 utilizing perception system 402 are illustrated. In some embodiments, CNN440 (e.g., one or more components of CNN 440) is the same as or similar to CNN 420 (e.g., one or more components of CNN 420) (see fig. 4B).

At step 450, perception system 402 provides data associated with the image as input to CNN440 (step 450). For example, as illustrated, the perception system 402 provides data associated with an image to the CNN440, where the image is a grayscale image represented as values stored in a two-dimensional (2D) array. In some embodiments, the data associated with the image may include data associated with a color image represented as values stored in a three-dimensional (3D) array. Additionally or alternatively, the data associated with the image may include data associated with an infrared image and/or a radar image, or the like.

At step 455, cnn440 performs a first convolution function. For example, based on CNN440 providing a value representing an image as input to one or more neurons (not explicitly illustrated) included in first convolution layer 442, CNN440 performs a first convolution function. In this example, the value representing the image may correspond to the value of the region (sometimes referred to as receptive field) representing the image. In some embodiments, each neuron is associated with a filter (not explicitly illustrated). The filter (sometimes referred to as a kernel) may be represented as an array of values corresponding in size to the values provided as inputs to the neurons. In one example, the filter may be configured to identify edges (e.g., horizontal lines, vertical lines, and/or straight lines, etc.). In successive convolutional layers, filters associated with neurons may be configured to continuously identify more complex patterns (e.g., arcs and/or objects, etc.).

In some embodiments, CNN 440 performs a first convolution function based on CNN 440 multiplying the value provided as input to each of the one or more neurons included in first convolution layer 442 with the value of the filter corresponding to each of the same one or more neurons. For example, CNN 440 may multiply the value provided as an input to each of the one or more neurons included in first convolutional layer 442 by the value of the filter corresponding to each of the one or more neurons to generate a single value or array of values as an output. In some embodiments, the collective outputs of the neurons of the first convolutional layer 442 are referred to as convolutional outputs. In some embodiments, where the individual neurons have the same filter, the convolved output is referred to as a signature.

In some embodiments, CNN 440 provides the output of each neuron of first convolutional layer 442 to neurons of a downstream layer. For clarity, the upstream layer may be a layer that transfers data to a different layer (referred to as a downstream layer). For example, CNN 440 may provide the output of each neuron of first convolutional layer 442 to a corresponding neuron of the sub-sampling layer. In an example, CNN 440 provides the output of each neuron of first convolutional layer 442 to a corresponding neuron of first sub-sampling layer 444. In some embodiments, CNN 440 adds bias values to the set of all values provided to the various neurons of the downstream layer. For example, CNN 440 adds bias values to the set of all values provided to the individual neurons of first sub-sampling layer 444. In such an example, CNN 440 determines the final value to be provided to each neuron of first sub-sampling layer 444 based on the set of all values provided to each neuron and the activation function associated with each neuron of first sub-sampling layer 444.

At step 460, cnn440 performs a first sub-sampling function. For example, CNN440 may perform a first sub-sampling function based on CNN440 providing the values output by first convolutional layer 442 to the corresponding neurons of first sub-sampling layer 444. In some embodiments, CNN440 performs a first sub-sampling function based on the aggregate function. In an example, the CNN440 performs a first sub-sampling function based on the CNN440 determining the largest input (referred to as the max-pooling function) of the values provided to a given neuron. In another example, the CNN440 performs a first sub-sampling function based on the CNN440 determining an average input (referred to as an average pooling function) in the values provided to a given neuron. In some embodiments, based on CNN440 providing values to the various neurons of first sub-sampling layer 444, CNN440 generates an output, sometimes referred to as a sub-sampled convolutional output.

At step 465, cnn440 performs a second convolution function. In some embodiments, CNN440 performs a second convolution function in a manner similar to how CNN440 performs the first convolution function described above. In some embodiments, CNN440 performs a second convolution function based on CNN440 providing as input the value output by first sub-sampling layer 444 to one or more neurons (not explicitly illustrated) included in second convolution layer 446. In some embodiments, as described above, each neuron of the second convolution layer 446 is associated with a filter. As described above, the filter(s) associated with the second convolution layer 446 may be configured to identify more complex patterns than the filter associated with the first convolution layer 442.

In some embodiments, CNN 440 performs a second convolution function based on CNN 440 multiplying the value provided as input to each of the one or more neurons included in second convolution layer 446 with the value of the filter corresponding to each of the one or more neurons. For example, CNN 440 may multiply the value provided as an input to each of the one or more neurons included in second convolution layer 446 with the value of the filter corresponding to each of the one or more neurons to generate a single value or array of values as an output.

In some embodiments, CNN 440 provides the output of each neuron of second convolutional layer 446 to neurons of a downstream layer. For example, CNN 440 may provide the output of each neuron of first convolutional layer 442 to a corresponding neuron of the sub-sampling layer. In an example, CNN 440 provides the output of each neuron of first convolutional layer 442 to a corresponding neuron of second sub-sampling layer 448. In some embodiments, CNN 440 adds bias values to the set of all values provided to the various neurons of the downstream layer. For example, CNN 440 adds bias values to the set of all values provided to the individual neurons of second sub-sampling layer 448. In such an example, CNN 440 determines the final value provided to each neuron of second sub-sampling layer 448 based on the set of all values provided to each neuron and the activation function associated with each neuron of second sub-sampling layer 448.

At step 470, cnn440 performs a second sub-sampling function. For example, CNN440 may perform a second sub-sampling function based on CNN440 providing the values output by second convolution layer 446 to corresponding neurons of second sub-sampling layer 448. In some embodiments, CNN440 performs a second sub-sampling function based on CNN440 using an aggregation function. In an example, as described above, CNN440 performs a first sub-sampling function based on CNN440 determining the maximum or average input of the values provided to a given neuron. In some embodiments, CNN440 generates an output based on CNN440 providing values to individual neurons of second sub-sampling layer 448.

At step 475, cnn440 provides the output of each neuron of second sub-sampling layer 448 to full connection layer 449. For example, CNN440 provides the output of each neuron of second sub-sampling layer 448 to fully connected layer 449, such that fully connected layer 449 generates an output. In some embodiments, the fully connected layer 449 is configured to generate an output associated with the prediction (sometimes referred to as classification). The prediction may include an indication that an object included in the image provided as input to CNN440 includes an object and/or a collection of objects, etc. In some embodiments, the perception system 402 performs one or more operations and/or provides data associated with predictions to the different systems described herein.

Lane classification system

As noted above, autonomous vehicles typically rely largely on maps to accurately reflect their current operating environment. For example, safe and efficient operation of AV may depend to a large extent on knowledge of the current roadway, such as where the boundaries of the roadway are and the configuration of lanes within the roadway. The configuration is typically marked by various lane markings such as lane separation lines (dividing lanes in the same direction), road separation lines (dividing lanes in different directions), and road boundaries. While these identifications may be visually depicted in a camera image, it may be difficult to capture the exact location of these identifications using such an image. Furthermore, these identifications may be difficult to detect programmatically in more accurate sensor data (such as lidar data, etc.).

Embodiments of the present invention address these difficulties by providing a lane classification system 504, where the lane classification system 504 may accurately detect lane identifications within laser radar data or other 3D imaging data and use such identifications to classify lanes of a roadway. In one embodiment, lane identification and/or lane classification is used to control the operation of the AV. In another embodiment, lane identification and/or lane classification is used to generate highly accurate maps of the roadway, which may be loaded into other AV's (e.g., lacking a lidar sensor modality) to facilitate operation of these other AV's, for example. In one embodiment, the lane classification system 504 is implemented by a computing system within the AV (such as the apparatus 300 of fig. 3 included within the vehicle 102 of fig. 1, etc.). In another embodiment, the lane classification system 504 is implemented by a computing system external to the AV (such as the apparatus 300 of fig. 3 implemented independently of the vehicle 102 of fig. 1, etc.).

As shown in fig. 5, the lane classification system 504 illustrates operating based on the scene data 502. Scene data 502 illustratively includes data from a particular environment, such as a given roadway, etc. In one embodiment, the scene data 502 includes lidar data, which may correspond to a 3D representation of the environment (such as a point cloud). In other embodiments, the scene data 502 may additionally or alternatively include data from other sensor modalities, such as camera images or radar images, and the like. The scene data may include additional data related to the environment (but may not be captured from the environment), such as location information provided by GPS or other accurate positioning technology, and the like. In one example, the scene data 502 includes data captured from an environment via a single sensor modality. For example, scene data 502 (e.g., in addition to other data related to the environment but not captured from the environment (such as GPS sites, etc.) may include data captured from the environment via lidar sensors.

In some cases, the scene data 502 may be "real-time" data, such as a data stream reflecting the current state of the environment, and the like. In other cases, the scene data 502 may be previously captured and stored data, and thus may not reflect the current state of the environment. Further, the scene data 502 may reflect aggregated data from a single scan or multiple scans of the environment. Illustratively, the scene data 502 may correspond to an aggregation of multiple lidar scans of an area over time and potentially captured at different locations in the area. In some cases, such aggregate data may increase the accuracy of the scene data 502.

As shown in fig. 5, the scene data 502 is fed to a pseudo-image generating system 506, which pseudo-image generating system 506 is configured to generate a pseudo-image from the scene data 502. In general, the pseudo-image may correspond to a 2D representation of 3D data, such as lidar data. For example, the pseudo-image may "compress" multiple data points of the 3D data along a given dimension, such as a "column" of points along a given value of the height dimension (Z-axis), into a single data point in the resulting 2D data. The data points within the pseudo-image may be enhanced with information about the plurality of points from which the pseudo-image was created, such as maximum and minimum values of attributes of the plurality of points within a compressed dimension represented by the data points in the pseudo-image, or statistical aggregation of attribute values between the plurality of points, etc. Thus, the pseudo-image may enable processing of 3D data (such as lidar data, etc.) via 2D processing techniques. In one embodiment, the pseudo-image is generated according to the technique discussed in Alex H.Lang et al, "PointPicloras: fast Encoders for Object Detection from Point Clouds" (arXiv: 1812.05784, available at https:// doi.org/10.48550/arXiv.1812.05784), the entire contents of which are incorporated herein by reference. In some cases, the pseudo-image generating system 506 is implemented using one or more neural networks trained to receive 3D image data and output pseudo-images. In one example, the pseudo-image is a BEV image of the region.

In one example, the pseudo-image generating system 506 generates a pseudo-image from the scene data 502 representing one scan of the environment. In another example, the pseudo-image generating system 506 generates a pseudo-image from the scene data 502 representing multiple scans of the environment. Illustratively, the pseudo-image generation system 506 may aggregate multiple scans into a single aggregate lidar point cloud, such as by overlaying the data of each scan when adjusting the differences between the scans (e.g., by adjusting the relative positions of the lidar sensors between the scans). Additionally or alternatively, the scene data 502 may include such an aggregated point cloud. Further, in some embodiments, the lane classification system 504 may be configured to directly accept a pseudo-image or other 2D imagery, such that the pseudo-image generation system 506 may be omitted or the pseudo-image generation system 506 may not be utilized for such imagery.

In fig. 5, the pseudo-image or other 2D image is then passed to a road element classification system 508, which road element classification system 508 illustratively implements a neural network machine learning architecture to identify pixels within the pseudo-image or other 2D image that correspond to particular road elements (such as lane separation lines, road separation lines, or road boundaries, etc.). As described in more detail below, the road element classification system 508 may implement CNN in conjunction with a sub-network of scene feature vectors that generate input data (e.g., pseudo-images). The scene feature vectors may be used to capture context information related to the entire scene data 502, facilitating differentiation between, for example, road and lane separation lines, which may otherwise be difficult to differentiate. In one example, the scene feature vector is used as an input during a deconvolution layer of the machine learning architecture such that the output of the machine learning architecture contains features learned from the scene feature vector, thereby enabling more accurate detection of road features. Illustratively, the road element classification system 508 may provide as output a 2D image with a label, where the label indicates, for example, whether each pixel of the input 2D image corresponds to a given road element.

The output of the road element classification system 508 is then passed to the lane instance system 510. In fig. 5, the lane instance system 510 is configured to generate, from the annotated 2D image, a set of multi-segment lines representing lanes within an area corresponding to the 2D image. For example, the lane instance system 510 may accept 2D images with pixels annotated with confidence values corresponding to various potential road elements. As described in more detail below, the lane instance system 510 may apply various transforms, such as thresholding and skeletonize(s), to the 2D image to identify a set of pixels predicted to correspond to a particular road element. The lane instance system 510 may then convert the set of pixels into a multi-segment line (a continuous line representing one or more line segments of the road element). Thus, a 2D image containing pixels associated with confidence values may be converted into a set of edges and vertices representing road elements.

The multi-segment lines generated by the lane instance system 510 may then be used as a highly accurate representation of road features within the region corresponding to the entered scene data 502. For example, the multi-segment line may be passed to a mapping system 512, which mapping system 512 is configured to generate a map of the area. Illustratively, the mapping system 512 may be configured to combine the multi-segment lines with other information about the region to generate a region map, which may illustratively be transmitted to one or more AV's for navigation in the region.

As another example, the multi-segment line may be passed to planning system 514 to facilitate planning. For example, planning system 514 may correspond to planning system 404 of FIG. 4A. Thus, the planning system 514 may utilize the multi-segment lines to navigate through the area represented by the scene data 502, such as by planning a route to maintain a current lane, planning lane changes, and detecting other traffic.

To further illustrate the operation of the lane classification system 504, fig. 6 depicts an example image processed at various points in the process of fig. 5. As shown in fig. 6, the scene data 502 may correspond to multiple instances of lidar data 601 fed into the lane classification system 504. Examples may correspond to point clouds acquired at different locations in an area, for example.

The lane classification system 504 may then aggregate the lidar data instances 601 into an aggregate lidar data set 603 at 602. The aggregate lidar data set 603 may, for example, provide higher resolution data than the lidar data instance 601 alone. As described above, in some embodiments, lane classification system 504 may operate on individual lidar data instances 601 instead of aggregate lidar data sets 603.

Thereafter, the lane classification system 504 generates a pseudo image 605 at 604. As described above, the pseudo-image may reduce the dimensionality of the aggregate lidar data 603, such as by representing the 3D point cloud as a annotated 2D image, or the like. The individual pixels of the pseudo-image are illustratively labeled with data corresponding to voxels in the aggregate lidar data 603 corresponding to those pixels, thereby enabling the 3D lidar data 603 to be represented in fewer dimensions.

The lane classification system 504 may then pass the pseudo image 605 through the machine learning model to classify the respective pixels of the 2D image as corresponding to the respective road elements at 606. For example, the model may assign to each pixel a probability that the pixel corresponds to a lane separation line, a probability that the pixel corresponds to a road separation line, and a probability that the pixel corresponds to a road boundary. As discussed in more detail below, the machine learning model may correspond to a CNN having a sub-network that generates scene feature vectors of the input data from an intermediate representation generated by the CNN, thereby enabling the model to capture both global and local information within the input data and provide a more accurate classification of individual pixels corresponding to particular road elements.

As a result of the classification at 606, the lane classification system 504 may obtain information (e.g., metadata) indicating the probability that each pixel within the pseudo-image 605 corresponds to a given classification feature (such as a road element of a given type, etc.). In one embodiment, the lane classification system 504 applies a threshold (which may be set during training based on a empirical review of the results) to remove pixels that do not correspond to any classification feature. For example, the lane classification system 504 may remove pixels (e.g., 30%, 50%, 70%, 90%, 95%, etc.) from the pseudo-image 605 that do not have a classification above the threshold. The lane classification system 504 may also assign a classification to each pixel that corresponds to the feature with the highest probability. For example, if a pixel is assigned a probability of 50% as a lane separation line and a probability of 60% as a road separation line, the lane classification system 504 may assign a classification as a road separation line to the pixel. As a result, the lane classification system 504 may generate a classification image 607, the classification image 607 comprising pixels assigned to different classes according to the classification at 606. For example, in fig. 6, the outermost line of the image 607 may represent a detected road boundary, the 5 th line from the left (shown in dotted line in fig. 6) may represent a detected road separation line, and the remaining lines may represent detected lane separation lines. Since the pseudo-image 605 is associated with highly accurate location information, the locations of these pixels can be converted to physical locations in the area corresponding to the pseudo-image 605 with high accuracy. Therefore, the AV navigating in the area can be notified of the physical location of each road element with high confidence.

In some cases, it may be beneficial to further process the pixelated image (such as classification image 607, etc.) for use by a computing device (such as a computing device within an AV, etc.). For example, it may be helpful to convert pixel values that may be difficult to map to physical locations alone or to use in computation into multi-segment lines. As used herein, a multi-segment line generally refers to a continuous line of one or more line segments (which may be represented, for example, as vertices and a collection of edges between the vertices). The multi-segment line may be more efficiently stored (as many pixels may be represented by one edge, for example) and more easily used in computation than a pixel representation. Thus, in some embodiments, the classified image 607 may be used to generate a set of multi-segment lines at 608. In general, the conversion to a multi-segment line may include applying one or more morphological operations to the classified image 607 and converting the deformed image to a multi-segment line via a multi-segment line conversion algorithm. In one embodiment, the conversion to the multi-segment lines occurs simultaneously for all types of classifications (e.g., all types of road elements), and then the multi-segment lines so obtained are associated with classifications based on the classifications of the pixels from which they were created. In another embodiment, the transformations to the multi-segment lines occur independently for each type of classification, with the multi-segment lines created at each transformation being marked as corresponding to the relevant classification for that transformation. Illustratively, the conversion may occur by creating a binary image from a classified image 607 that includes only pixels having a relevant classification above a given threshold. The converting may further include: the binary image is skeletonized, such as by thinning all lines within the image to a given width (e.g., one pixel), and so forth. Additionally, converting may include applying a multi-segment line conversion algorithm to convert pixels (e.g., of the skeletonized binary image) into multi-segment lines. In one embodiment, the conversion algorithm includes: each set of connected pixels in the skeletonized binary image is considered as an undirected graph, and the longest path of no loop or branch in the undirected graph is found. Thus, the set of connected pixels can be converted into vertices and edges corresponding to the multi-segment lines. The multi-segment line may in turn be represented within a map of the area from which the obtained scene data originated, thereby enabling a programmed detection of physical road elements.

Referring to fig. 7, an exemplary machine learning architecture 700 for detecting features within a 2D image using CNNs and global feature subnets will be described. The architecture may be implemented by the lane classification system 504, for example, as a road element classification system 508.

Architecture 700 includes both CNN 702 and subnetwork 704 portions. The CNN 702 section illustratively includes a series of convolutional layers followed by a series of deconvolution layers. For example, CNN 702 may be implemented as a U-net network, where the convolutional layers form a contracted path and the deconvolution layers form an expanded path. The convolutional layer may be implemented, for example, as described in connection with fig. 4C and 4D. The deconvolution layer may generally operate in the opposite manner as the convolution layer, up-sampling and up-converting the smaller input to the larger intermediate output. In one embodiment, the convolutional layer and the deconvolution layer are symmetrical such that the final deconvolution layer provides the output or the same dimension as the input. As shown in fig. 7, each deconvolution layer illustratively passes as input a set of features from a previous layer (e.g., a final deconvolution layer or a previous deconvolution layer) and data of the corresponding deconvolution layer (e.g., where the output of a first deconvolution layer is passed as input to the final deconvolution layer, the output of a second deconvolution layer is passed as input to the penultimate deconvolution layer, etc.). In one embodiment, inputs to the deconvolution layers are concatenated, such as by treating each input as one or more channels of a multi-channel input, or the like. The output of the final deconvolution layer may then be provided to a regression head (regression head) that classifies each pixel of the output as corresponding to one or more road features. For example, the regression head may assign a plurality of classification values to each pixel, each classification value indicating the likelihood that the pixel corresponds to a given classification.

As shown in fig. 7, each deconvolution layer of architecture 700 takes as input the output of sub-network 704 in addition to the output of the previous layer and the corresponding convolution layer. In fig. 7, the output of the subnetwork 704 is a scene feature vector that provides global context (e.g., a pseudo-image representing a high accuracy laser radar scan of an area) related to the input to the architecture 700, facilitating contextual classification of features, such as differentiation between road elements within the input that may appear to be surface-similar, etc. Specifically, in fig. 7, the sub-network 704 includes a multi-layer perceptron, sometimes referred to as a fully-connected feed-forward neural network, that takes as input data corresponding to at least one intermediate representation generated by the convolutional layer of the CNN 702. The input data illustratively has a reduced dimension relative to the intermediate representation. For example, each 2D intermediate representation may be processed via a pooling operation, such as average pooling, to produce a vector. These vectors may be concatenated together to provide an input to the MLP. In fig. 7, vectors corresponding to the convolutional layers of CNN 702 are shown; however, in some embodiments, the MLP may take as input vectors from less than all of the convolutional layers of CNN 702. The output of the MLP is illustratively a scene feature vector. The output may then be provided as an input to one or more deconvolution layers. To account for dimensional differences between the output of the MLP and the input to each deconvolution layer, the subnetwork 704 may appropriately tile the output of the MLP for each deconvolution layer, such as by repeating the output of the MLP for each input location of the deconvolution layer. Since the output is generated based on an intermediate representation of the convolutional layer, the output may provide a global context related to the input to the CNN 702. Thus, the output of the MLP in the subnet 704 may enable the deconvolution layer to take into account the global context of the input to the CNN 702. This in turn increases the accuracy of the classification produced by architecture 700, thereby enabling, for example, improved accuracy in classifying road elements within a pseudo-image corresponding to a lidar scan of the environment surrounding the AV.

Architecture 700 of fig. 7 may be implemented during both training and inference. For example, training of architecture 700 may occur based on labeled pseudo-images, wherein individual pixels of the labeled image are labeled as corresponding to particular road features. Thus, training may occur through the following operations: the marked image is passed through the architecture 700 and weights or operations within the architecture 700 are adjusted so that differences between the output of the architecture 700 and the marked data are minimized. The unlabeled pseudo-image may then be passed through a trained architecture 700 (e.g., architecture 700 with weights established during training) to generate an output.

Referring to fig. 8, an exemplary routine 800 for classifying road elements within lidar scan data to provide high accuracy detection of road elements within a physical region will be described. The routine 800 is illustratively implemented by the lane classification system 504 (e.g., as implemented on the device 300, possibly within the vehicle 200).

Routine 800 begins at block 802, where the lane classification system 504 obtains a pseudo image of an area corresponding to one or more lidar scans of the area. As discussed above, the pseudo-image may represent a 2D representation of the lidar scan, such as by converting the height dimension into annotations for each pixel in the 2D representation. The pseudo-image may be generated based on data from a single lidar scan, or based on an aggregate lidar image representing data from multiple lidar scans (e.g., where the data of each lidar scan is adjusted to account for positional differences between scans).

The routine 800 then proceeds to block 804, where the lane classification system 504 determines a road element classification for pixels within the pseudo image at block 804. As discussed above, the road element classification may correspond to any number of road elements (e.g., including lane separation lines, road separation lines, and road boundaries). In one embodiment, the lane classification system 504 generates the classification using the machine learning architecture 700 of fig. 7 as previously trained on a set of training data (e.g., labeled pseudo-images). As described above, architecture 700, by way of inclusion of a sub-network that considers global context via scene feature vectors, may be particularly suited to distinguishing superficially similar elements (such as lane and road separation lines, etc.) that are otherwise indistinguishable within a pseudo-image. Thus, in one embodiment, the classification determined at block 804 includes classifications of both lane and road separation lines, thereby enabling determination of traffic direction of lanes in the area.

At block 806, the lane classification system 504 generates a multi-segment line based on the pixel classification resulting from the implementation of block 804. Illustratively, the generating of the multi-segment line may include: the image is binarized, such as by removing (e.g., setting to a value of 0) any pixels having a correlation classification that does not satisfy the threshold, and equalizing (e.g., setting to a value of 1) all pixels having a correlation classification that satisfies the threshold. The generation of the multi-segment line may also include morphological operations such as skeletonization of images. Further, the generating of the multi-segment line may include: a multi-segment line approximation algorithm is applied to the image, such as by finding the longest non-circular path between adjacent positive pixels in the image and computing a set of edges and vertices representing the path. Each polyline may then be associated with metadata that indicates a classification with the pixel from which the polyline was created. In one embodiment, the transition to the multi-segment line occurs simultaneously for all types of classifications (e.g., all types of road elements). For example, each pixel may maintain metadata indicating a classification of the pixel, and one or more multi-segment lines may be generated for each classified pixel, wherein the one or more multi-segment lines are labeled with the classification. In another embodiment, the conversion to the multi-segment line occurs independently for each type of classification. For example, a binary image may be created for each class (e.g., where pixels that meet the threshold of the class are set to true or positive values and all other pixels are set to negative), after which a multi-segment line corresponding to the class may be created. The thus obtained classified multi-segment lines may then be combined to represent the relevant elements in the region.

As discussed above, since the pseudo-image may represent a high accuracy scan of the physical region, and since the lane classification system 504 provides a high accuracy classification of the scanned element, the generated multi-segment line may then be used as a high accuracy representation of the road element. For example, the classification may be provided to one or more AV's to provide these AV's with high accuracy location information of the road elements, thereby enabling safe and efficient programmed navigation of the region by the AV's.

Referring to fig. 9, an illustrative routine 900 for image segmentation based on a convolutional neural network and a global context sub-network will be described, which as described above may be used to make the classification as described with respect to fig. 8. Routine 900 is illustratively implemented by lane classification system 504 (e.g., as implemented on device 300, possibly within vehicle 200). For example, routine 900 (such as during inference, etc.) may be implemented as an implementation of architecture 700 of fig. 7 to detect and classify road elements reflected in data collected during one or more lidar scans.

The routine 900 begins at block 902, where the lane classification system 504 obtains an input 2D image. The image may be, for example, a pseudo image generated based on data corresponding to one or more lidar scans.

At block 904, the lane classification system 504 generates a first feature map from the input image. Illustratively, the first feature map may be generated by convolving the input image during a convolution layer (such as the layer described above with reference to the architecture of fig. 7, etc.). The feature map may capture features of the input image based on, for example, one or more kernels used to generate a first feature map, which may be a result of training of the architecture.

At block 906, the lane classification system 504 generates a first vector based on the first feature map. In general, the vector may be a transformation of the first feature map that reduces the dimension of the first feature map. For example, where the first feature map is a matrix of dimensions x y C (where x and y represent coordinates in the 2D plane and C represents the channel of pixels at these coordinates), the feature map may have dimension C, thereby compressing the data from all coordinates into a single dimension. In one embodiment, the vector is generated by applying an average pooling operation on the elements of the first feature map. In other embodiments, other pooling operations (e.g., max-pooling, min-pooling, etc.) may be used. Thus, the first vector may be considered to capture the global context from the feature map.

At block 908, the lane classification system 504 generates a second feature map of the input image. Illustratively, the second feature map may be generated by a second or subsequent convolution layer of an architecture (such as architecture 700 of fig. 7). Thus, the second feature map may be generated by a direct convolution of the first feature map, by a convolution of an intermediate feature map generated from a convolution of the first feature map, by multiple convolutions of such intermediate feature map, or the like. Thus, the second feature map may be considered to capture "higher level" features of the input image, as the second feature map is generated by more convolutions relative to the first feature map.

At block 910, the lane classification system 504 generates a second vector based on the second feature map. As described above, a vector may be a transformation of a corresponding feature map that reduces the dimensions of the feature map. In one embodiment, the second vector, when applied to the second feature map, is generated by the same operation as the first vector. For example, the two vectors may be generated by applying a given pooling operation to their respective graphs. In another embodiment, the second vector is generated by a different operation than the first vector (such as a different pooling operation, etc.). For example, the second vector may be generated by applying a maximum pooling operation to the second feature map, while the first vector may be generated by applying an average pooling operation to the first feature map. As with the first vector, the second vector may be considered to capture the global context from the second feature map because the second vector has a reduced dimension relative to the second feature map.

At block 912, the lane classification system 504 generates a scene feature vector based on the first vector and the second vector. For example, the lane classification system 504 may pass a cascade of first vectors and second vectors through a multi-layer perceptron that outputs scene feature vectors. The multi-layer perceptron may be trained to extract features related to the ongoing classification task from the first vector and the second vector. For example, the MLP may be trained to extract features within the first and second vectors that relate to classification of road elements (e.g., as road dividers, lane dividers, boundaries, etc.). When the vector captures the global context of each respective feature map, the scene feature vector may similarly be considered to capture the global context across the vector and thus across the convolution of the input image.

Although fig. 9 depicts generating a scene feature vector based on the first vector and the second vector, in some embodiments, a scene feature vector may be generated based on other vectors. For example, the lane classification system 504 may generate feature vectors for each convolutional layer and generate scene feature vectors based on a concatenation of all such feature vectors. Thus, an architecture with n convolutional layers may generate scene feature vectors from concatenating n feature vectors (each corresponding to a feature map generated by one of the convolutional layers) through the MLP.

At block 914, the lane classification system 504 generates an image segmentation map based on the scene feature vector and the second feature map. Illustratively, the second feature map may be output by a final convolution layer within the machine learning architecture, and the lane classification system 504 may generate the image segmentation by deconvolving a combination (e.g., concatenation) of the second feature map and the scene feature vector. Since the scene feature vector illustratively has a lower dimension than the second feature map, the lane classification system 504 may tile the scene feature vector during deconvolution, such as by cascading the scene feature vector to pixels of the second feature map.

In some embodiments, the image segmentation map may be generated based on directly applying the scene feature vector to the second feature map, such as by deconvolving a concatenation of the second feature map and the scene feature vector. In other embodiments, the image segmentation map may be generated based on applying the scene feature vector to data generated from the second feature map. For example, the second feature map may be deconvolved one or more times, and the second feature vector may be concatenated with the results of these deconvolution operations and passed through a final deconvolution operation to produce an image segmentation map. In yet another embodiment, the architecture may concatenate the scene feature vectors to each given data set prior to convolving the given data set. For example, the architecture may concatenate the scene feature map to the second feature map prior to initial deconvolution, concatenate the scene feature map (e.g., with appropriate tiling) to the deconvolution's output prior to subsequent deconvolution, and so on. The final result of the deconvolution is illustratively passed through a regression head to generate an image segmentation map. As discussed above, since the lane classification system 504 considers scene feature vectors and feature maps during the generation of the image segmentation map, the accuracy of the segmentation map may be improved relative to other methods.

Although fig. 9 depicts generating an image segmentation map based on the scene feature vector and the second feature map, the lane classification system 504 may additionally or alternatively output other maps. For example, the lane classification system 504 may generate a Distance Transformation (DT) graph (such as an inverse DT graph, etc.) that indicates the likelihood of each pixel corresponding to a road element as the relative distance of that pixel relative to the nearest instance of the road element. Illustratively, different inverse DT maps may be generated for each type of road element.

Various example embodiments of the invention may be illustrated by the following clauses:

clause 1. A computer-implemented method, implemented by at least one processor, the computer-implemented method comprising:

receiving, with the at least one processor, an image representing lidar scanned data from a vehicle environment;

convolving, with the at least one processor, the image to generate a first feature map;

converting, with the at least one processor, the first feature map into a first feature vector;

applying, with the at least one processor, one or more additional convolutions to the first feature map to generate a second feature map;

Converting, with the at least one processor, the second feature map into a second feature vector;

passing, with the at least one processor, an input representing at least the first feature vector and the second feature vector through a neural network to generate a scene feature vector; and

an output map is generated, with the at least one processor, based at least on the scene feature vector and the second feature map, wherein the output map includes a plurality of pixels and indicates a likelihood that each pixel corresponds to a road element.

Clause 2. The computer-implemented method of clause 1, wherein the image is a two-dimensional pseudo image generated from a three-dimensional lidar point cloud image.

Clause 3 the computer-implemented method of clause 2, wherein the three-dimensional lidar point cloud image is generated from an aggregate three-dimensional lidar point cloud image of multiple lidar scans representing the vehicle environment.

Clause 4 the computer-implemented method of any of clauses 1 to 3, wherein generating the output map comprises: a segmentation map is generated that associates a classification value with each of the plurality of pixels, the classification value indicating a likelihood that the pixel corresponds to the road element.

Clause 5 the computer-implemented method of clause 4, wherein the road element is one of a plurality of road elements, and wherein the output map associates a plurality of classification values with each pixel, each classification value of the plurality of classification values corresponding to a corresponding road element of the plurality of road elements.

Clause 6 the computer-implemented method of clause 5, wherein the plurality of road elements comprises at least one of: lane separation lines indicating a division between traffic lanes in the same direction, road separation lines indicating a division between traffic lanes in different directions, road boundary elements indicating a boundary of a roadway, a parking line, a pedestrian sign, a bicycle sign, and a V-shaped sign.

Clause 7 the computer-implemented method of any of clauses 1 to 6, wherein applying one or more additional convolutions to the first feature map to generate the second feature map comprises: a first additional convolution is applied to the first feature map to generate an intermediate feature map and a second additional convolution is applied to the intermediate feature map to generate the second feature map.

The computer-implemented method of any of clauses 1-7, wherein converting the first feature map into the first feature vector comprises: a pooling operation is used to reduce the dimension of the first feature map.

Clause 9 the computer-implemented method of clause 8, wherein the pooling operation is an average pooling operation that reduces the first feature map to a single dimension.

The computer-implemented method of any of clauses 1-9, wherein the input representing at least the first feature vector and the second feature vector is a concatenation of at least the first feature vector and the second feature vector.

Clause 11 the computer-implemented method of any of clauses 1 to 10, wherein the one or more additional convolutions further generate additional feature maps, and wherein the input representing at least the first feature vector and the second feature vector further represents the additional feature maps.

The computer-implemented method of any of clauses 1-11, wherein generating the output map based on the scene feature vector and the second feature map comprises: deconvolving the second feature map based at least on the scene feature vector.

Clause 13 the computer-implemented method of clause 12, wherein deconvolving the second feature map is further based on the first feature map.

Clause 14 the computer-implemented method of clause 12, wherein deconvolving the second feature map based at least on the scene feature vector comprises: tiling the scene feature vectors.

Clause 15 the computer-implemented method of clause 12, wherein generating the output map based on the second feature vector and the second feature map further comprises: and applying a regression head to the deconvolution result of the second feature map.

Clause 16 the computer-implemented method of any of clauses 1 to 14, further comprising: one or more multi-segment lines are generated from the output map, each of the one or more multi-segment lines being used to identify road elements in the vehicle environment.

Clause 17 the computer-implemented method of clause 16, wherein generating the plurality of multi-segment lines comprises: and skeletonizing the output graph.

Clause 18 the computer-implemented method of clause 16, wherein generating the plurality of multi-segment lines comprises:

Converting one or more adjacent pixels in the output map having classification values meeting a threshold into an undirected map; and

the longest path within the undirected graph is identified as a multi-segment line of the one or more multi-segment lines.

Clause 19, a system comprising:

one or more non-transitory data stores comprising computer-executable instructions; and

one or more hardware processors configured to execute the computer-executable instructions to:

receiving an image representing data from a lidar scan of a vehicle environment;

convolving the image to generate a first feature map;

converting the first feature map into a first feature vector;

applying one or more additional convolutions to the first feature map to generate a second feature map;

converting the second feature map into a second feature vector;

passing an input representing at least the first feature vector and the second feature vector through a neural network to produce a scene feature vector; and

an output graph is generated based at least on the scene feature vector and the second feature map, wherein the output graph includes a plurality of pixels and indicates a likelihood that each pixel corresponds to a road element.

Clause 20, one or more non-transitory computer-readable media storing computer-executable instructions that, when executed by a computing system comprising a hardware processor, cause the computing system to:

convolving the image to generate a first feature map;

converting the first feature map into a first feature vector;

converting the second feature map into a second feature vector;

Various additional example embodiments of the invention may be described by the following additional clauses:

clause 1. A computer-implemented method comprising:

Receiving, with at least one processor, data corresponding to an image of a lidar scan representing a vehicle environment;

determining, with the at least one processor and using a machine learning model, a road element classification for a plurality of pixels of the image, wherein the classification for a particular pixel of the plurality of pixels indicates a traffic direction of a traffic lane associated with the particular pixel of the image; and

a plurality of multi-segment lines are generated based on the classification for the plurality of pixels, wherein at least two of the plurality of multi-segment lines indicate a boundary between at least two traffic lanes for traffic traveling in different directions, with the at least one processor.

Clause 2 the computer-implemented method of clause 1, wherein receiving the data corresponding to the image comprises: data corresponding to a two-dimensional pseudo image generated from a three-dimensional lidar point cloud image is received.

Clause 3 the computer-implemented method of clause 2, wherein the two-dimensional pseudo-image includes annotations for each pixel reflecting a height dimension of the three-dimensional lidar point cloud image.

Clause 4. The computer-implemented method of clause 2 or 3, wherein the three-dimensional lidar point cloud image is generated from an aggregate three-dimensional lidar point cloud image representing multiple lidar scans of the vehicle environment.

Clause 5 the computer-implemented method of clause 4, wherein one or more of the multiple lidar scans of the vehicle environment are adjusted based on movement of a lidar sensor between the multiple lidar scans.

Clause 6 the computer-implemented method of any of the preceding clauses, wherein the at least one processor is included within an autonomous vehicle in the vehicle environment, and wherein the method is implemented by the autonomous vehicle.

Clause 7 the computer-implemented method of any of the preceding clauses, wherein the lidar scanning is by a lidar sensor on a vehicle, and wherein the at least one processor is remote from the vehicle.

The computer-implemented method of any of the preceding clauses, wherein determining the road element classification for the plurality of pixels of the image representing data from the lidar scan comprises: an output graph is generated that includes a plurality of pixels and that indicates a likelihood that each pixel corresponds to a road element of a plurality of road elements.

Clause 9 the computer-implemented method of clause 8, wherein the plurality of road elements comprises at least one of: lane separation lines indicating a division between traffic lanes in the same direction, road separation lines indicating a division between traffic lanes in different directions, road boundary elements indicating a boundary of a roadway, a parking line, a pedestrian sign, a bicycle sign, and a V-shaped sign.

The computer-implemented method of any of clauses 8 and 9, wherein generating the plurality of multi-segment lines comprises: and skeletonizing the output graph.

Clause 11 the computer-implemented method of any of clauses 8 to 10, wherein generating the plurality of multi-segment lines comprises:

Clause 12, a system comprising:

one or more processors configured to execute the computer-executable instructions to:

Receiving data corresponding to an image of a lidar scan representing a vehicle environment;

determining a road element classification for a plurality of pixels of the image, wherein the classification for a particular pixel of the plurality of pixels indicates a traffic direction of a traffic lane associated with the particular pixel of the image; and

a plurality of multi-segment lines is generated based on the classification for the plurality of pixels, wherein at least two of the plurality of multi-segment lines indicate a boundary between at least two traffic lanes for traffic traveling in different directions.

Clause 13 the system of clause 12, wherein the data corresponding to the image comprises data corresponding to a two-dimensional pseudo image generated from a three-dimensional lidar point cloud image.

Clause 14 the system of clause 13, wherein the three-dimensional lidar point cloud image is generated from an aggregate three-dimensional lidar point cloud image of multiple lidar scans representing the vehicle environment.

The system of any of clauses 12-14, wherein the one or more processors are included within an autonomous vehicle in the vehicle environment.

The system of any of clauses 12-15, wherein to determine the road element classification for the plurality of pixels of the image representing data from the lidar scan, the computer-executable instructions cause the system to generate an output graph comprising a plurality of pixels and indicating a likelihood that each pixel corresponds to a road element of a plurality of road elements.

Clause 17 the system of clause 16, wherein to generate the plurality of multi-segment lines, the computer-executable instructions cause the computing system to:

Clause 18, one or more non-transitory computer-readable media storing computer-executable instructions that, when executed by a computing system comprising a processor, cause the computing system to:

Clause 19, one or more non-transitory computer-readable media of clause 18, wherein to determine the road element classification for the plurality of pixels of the image representing data from the lidar scan, the computer-executable instructions cause the system to generate an output graph comprising a plurality of pixels and indicating a likelihood that each pixel corresponds to a road element of a plurality of road elements.

Clause 20 the one or more non-transitory computer-readable media of clause 19, wherein, to generate the plurality of multi-segment lines, the computer-executable instructions cause the computing system to:

In the foregoing specification, aspects and embodiments of the disclosure have been described with reference to numerous specific details that may vary from implementation to implementation. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense. The sole and exclusive indicator of the scope of the application, and what is intended by the applicants to be the scope of the application, is the literal and equivalent scope of the claims, including any subsequent amendments, that are issued from the present application in the specific form of the claims. Any definitions expressly set forth herein for terms contained in such claims shall govern the meaning of such terms as used in the claims. In addition, when the term "further comprises" is used in the preceding description or the appended claims, the phrase may be followed by additional steps or entities, or sub-steps/sub-entities of the previously described steps or entities.

Claims

1. A computer-implemented method implemented by at least one processor, the computer-implemented method comprising:

2. The computer-implemented method of claim 1, wherein the image is a two-dimensional pseudo image generated from a three-dimensional lidar point cloud image.

3. The computer-implemented method of claim 2, wherein the three-dimensional lidar point cloud image is generated from an aggregate three-dimensional lidar point cloud image of multiple lidar scans representing the vehicle environment.

4. The computer-implemented method of any of claims 1-3, wherein generating the output map comprises: a segmentation map is generated that associates a classification value with each of the plurality of pixels, the classification value indicating a likelihood that the pixel corresponds to the road element.

5. The computer-implemented method of claim 4, wherein the road element is one of a plurality of road elements, and wherein the output map associates a plurality of classification values with each pixel, each classification value of the plurality of classification values corresponding to a corresponding road element of the plurality of road elements.

6. The computer-implemented method of claim 5, wherein the plurality of road elements comprises at least one of: lane separation lines indicating a division between traffic lanes in the same direction, road separation lines indicating a division between traffic lanes in different directions, road boundary elements indicating a boundary of a roadway, a parking line, a pedestrian sign, a bicycle sign, and a V-shaped sign.

7. The computer-implemented method of any of claims 1 to 6, wherein applying one or more additional convolutions to the first feature map to generate the second feature map comprises: a first additional convolution is applied to the first feature map to generate an intermediate feature map and a second additional convolution is applied to the intermediate feature map to generate the second feature map.

8. The computer-implemented method of any of claims 1-7, wherein converting the first feature map into the first feature vector comprises: a pooling operation is used to reduce the dimension of the first feature map.

9. The computer-implemented method of any of claims 1 to 8, wherein the input representing at least the first feature vector and the second feature vector is a concatenation of at least the first feature vector and the second feature vector.

10. The computer-implemented method of any of claims 1-9, wherein generating the output map based on the scene feature vector and the second feature map comprises: deconvolving the second feature map based at least on the scene feature vector.

11. The computer-implemented method of any of claims 1 to 10, further comprising: one or more multi-segment lines are generated from the output map, each of the one or more multi-segment lines being used to identify road elements in the vehicle environment.

12. A system for a vehicle, comprising:

convolving the image to generate a first feature map;

converting the first feature map into a first feature vector;

converting the second feature map into a second feature vector;

13. The system of claim 12, wherein the computer-executable instructions, when executed by the one or more hardware processors, further cause the system to generate one or more multi-segment lines from the output map, each of the one or more multi-segment lines to identify road elements in the vehicle environment.

14. The system of claim 13, wherein to generate the one or more multi-segment lines, the instructions cause the system to skeletonize the output graph.

15. The system of claim 13, wherein to generate the one or more multi-segment lines, the instructions cause the system to:

16. A computer-implemented method, comprising:

17. The computer-implemented method of claim 16, wherein determining the road element classification for the plurality of pixels of the image representing data from the lidar scan comprises: an output graph is generated that includes a plurality of pixels and that indicates a likelihood that each pixel corresponds to a road element of a plurality of road elements.

18. The computer-implemented method of claim 17, wherein generating the plurality of multi-segment lines comprises: and skeletonizing the output graph.

19. The computer-implemented method of any of claims 17 to 18, wherein generating the plurality of multi-segment lines comprises:

the longest path within the undirected graph is identified as a multi-segment line of the plurality of multi-segment lines.

20. The computer-implemented method of any of claims 16-19, wherein the at least one processor is included within an autonomous vehicle in the vehicle environment, and wherein the method is implemented by the autonomous vehicle.