WO2022133911A1 - 目标检测方法、装置、可移动平台及计算机可读存储介质 - Google Patents

目标检测方法、装置、可移动平台及计算机可读存储介质 Download PDF

Info

Publication number
WO2022133911A1
WO2022133911A1 PCT/CN2020/139043 CN2020139043W WO2022133911A1 WO 2022133911 A1 WO2022133911 A1 WO 2022133911A1 CN 2020139043 W CN2020139043 W CN 2020139043W WO 2022133911 A1 WO2022133911 A1 WO 2022133911A1
Authority
WO
WIPO (PCT)
Prior art keywords
target detection
target
candidate
target object
detection information
Prior art date
Application number
PCT/CN2020/139043
Other languages
English (en)
French (fr)
Inventor
徐斌
Original Assignee
深圳市大疆创新科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市大疆创新科技有限公司 filed Critical 深圳市大疆创新科技有限公司
Priority to PCT/CN2020/139043 priority Critical patent/WO2022133911A1/zh
Publication of WO2022133911A1 publication Critical patent/WO2022133911A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present application relates to the technical field of target detection, and in particular, to a target detection method, a device, a movable platform, and a computer-readable storage medium.
  • the movable platform For the control of the movable platform, it is necessary to obtain the target detection results of the target object in the environment where the movable platform is located. Through the target detection results of the target object, the movable platform can be controlled to avoid obstacles in the moving direction of the movable platform, ensuring that the movable platform can move platform security.
  • the movable platform mainly controls the sensor to continuously collect the sensor data, and processes each frame of sensor data separately through the target detection algorithm to obtain the processing result of each frame of sensor data, and then transmits each frame of sensor data.
  • the processing results of the sensory data are fused, so that the target detection results of the target objects can be obtained.
  • processing each frame of sensor data separately cannot guarantee the stability and accuracy of the target detection results. Therefore, the stability and accuracy of the target detection results need to be improved.
  • embodiments of the present application provide a target detection method, device, movable platform, and computer-readable storage medium, which aim to improve the stability and accuracy of target detection results.
  • an embodiment of the present application provides a target detection method, including:
  • the first candidate area determine the second candidate area of the target object in the current frame sensor data
  • target detection information of the target object is determined.
  • an embodiment of the present application further provides a target detection device, where the target detection device includes a memory and a processor;
  • the memory is used to store computer programs
  • the processor is configured to execute the computer program and implement the following steps when executing the computer program:
  • the first candidate area determine the second candidate area of the target object in the current frame sensor data
  • target detection information of the target object is determined.
  • the embodiments of the present application also provide a movable platform, including:
  • a power system arranged on the platform body, for providing moving power for the movable platform
  • a sensor arranged on the platform body, for collecting sensing data
  • the target detection device is arranged in the platform body, and is used for determining the target detection information of the target object and also for controlling the movable platform.
  • an embodiment of the present application further provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor can implement the above-mentioned The steps of the object detection method.
  • the embodiments of the present application provide a target detection method, device, movable platform, and computer-readable storage medium.
  • the first candidate area is to determine the second candidate area of the target object in the sensor data of the current frame, and finally the target detection information of the target object is determined according to the second candidate area.
  • the first candidate area in the sensing data is determined, so the timing information between the sensing data of the current frame and the sensing data of the previous frame is considered, which greatly improves the stability and accuracy of the target detection result.
  • FIG. 1 is a schematic diagram of a scene for implementing the target detection method provided by the embodiment of the present application
  • FIG. 2 is a schematic diagram of another scenario for implementing the target detection method provided by the embodiment of the present application.
  • FIG. 3 is a schematic flowchart of steps of a target detection method provided by an embodiment of the present application.
  • FIG. 4 is a schematic diagram of a scenario in which a second candidate region is determined in an embodiment of the present application.
  • FIG. 5 is a schematic diagram of a scene of determining a target candidate region in an embodiment of the present application.
  • FIG. 6 is a schematic diagram of another scenario for determining a target candidate region in an embodiment of the present application.
  • Fig. 7 is the sub-step schematic flow chart of the target detection method in Fig. 3;
  • FIG. 8 is a schematic diagram of a scene of determining a third candidate region in an embodiment of the present application.
  • FIG. 9 is a schematic block diagram of the structure of a target detection apparatus provided by an embodiment of the present application.
  • FIG. 10 is a schematic structural block diagram of a movable platform provided by an embodiment of the present application.
  • the movable platform For the control of the movable platform, it is necessary to obtain the target detection results of the target object in the environment where the movable platform is located. Through the target detection results of the target object, the movable platform can be controlled to avoid obstacles in the moving direction of the movable platform, ensuring that the movable platform can move platform security.
  • the movable platform mainly controls the sensor to continuously collect the sensor data, and processes each frame of sensor data separately through the target detection algorithm to obtain the processing result of each frame of sensor data, and then transmits each frame of sensor data.
  • the processing results of the sensory data are fused, so that the target detection results of the target objects can be obtained.
  • each frame of sensor data is processed separately without considering the timing information between sensor data. Therefore, the stability and accuracy of target detection results need to be improved.
  • embodiments of the present application provide a target detection method, device, removable platform, and computer-readable storage medium.
  • acquiring the sensing data of the current frame and the first candidate region of the target object in the sensing data of the previous frame and then determining the second candidate region of the target object in the sensing data of the current frame according to the first candidate region, and finally according to the first candidate region
  • the second candidate area determines the target detection information of the target object. Since the second candidate area is determined based on the first candidate area of the target object in the sensor data of the previous frame, the sensor data of the current frame and the transmission data of the previous frame are considered.
  • the timing information between the sensory data greatly improves the stability and accuracy of the target detection results.
  • the object detection method can be applied to movable platforms, and the movable platforms include drones, robots, unmanned ships, and unmanned vehicles.
  • FIG. 1 is a schematic diagram of a scene for implementing the target detection method provided by the embodiment of the present application.
  • the driverless car 100 includes a car body 110 , a sensor 120 disposed on the car body 110 , and a power system 130 disposed on the car body 110 .
  • the sensor 120 is used to collect sensing data, and the power system 130 uses To provide mobility for the driverless car 100.
  • the sensor 120 includes a vision sensor and a radar device, the vision sensor may be a monocular vision sensor or a binocular vision sensor, and the radar device may include a lidar and a millimeter-wave radar.
  • driverless vehicle 100 may include one or more radar devices.
  • lidar can obtain laser point clouds by emitting laser beams to detect the position, speed and other information of objects in an environment.
  • Lidar can transmit detection signals to the environment including the target object, and then receive the reflected signal reflected from the target object, and obtain laser light according to the reflected detection signal, the received reflected signal, and data parameters such as the interval time between sending and receiving.
  • point cloud can include N points, and each point can include parameters such as x, y, z coordinates and intensity (reflectivity).
  • the unmanned vehicle 100 may further include a target detection device (not shown in FIG. 1 ), and the target detection device is used to obtain the current frame sensing data collected by the sensor 120 and the target object collected by the sensor 120 .
  • the first candidate area in the sensing data of the previous frame of the frame it is also used to determine the second candidate area of the target object in the sensor data of the current frame according to the first candidate area; it is also used to determine the target object according to the second candidate area Object detection information.
  • the target detection information includes the category, three-dimensional position coordinates, size, and confidence of the category of the target object.
  • FIG. 2 is a schematic diagram of another scenario for implementing the target detection method provided by the embodiment of the present application.
  • the UAV 200 includes a body 210, a sensor 220 provided on the body 210, and a power system 230 provided on the body 210.
  • the sensor 220 is used to collect sensing data
  • the power system 230 is used for the unmanned aerial vehicle.
  • the aircraft 200 provides flight power.
  • the sensor 220 includes a visual sensor and a radar device, and the radar device may include a lidar and a millimeter-wave radar.
  • drone 200 may include one or more radar devices.
  • one or more of the power systems 230 in the horizontal direction may rotate in a clockwise direction, and one or more of the power systems 230 in the horizontal direction may rotate in a counterclockwise direction.
  • the rotational rate of each power system 230 in the horizontal direction can be varied independently to achieve the lift and/or push operation caused by each power system 230 to adjust the spatial orientation, velocity and/or acceleration of the UAV 200 (eg, relative to up to three degrees of freedom for rotation and translation).
  • the power system 230 enables the drone 200 (the drone) to take off vertically from the ground, or to land vertically on the ground, without any horizontal movement of the drone 200 (the drone) ( if taxiing on the runway is not required).
  • the power system 230 may allow the drone 200 (drone) to pre-set positions and/or turn the steering wheel in the air.
  • One or more of the power systems 230 may be controlled independently of the other power systems 230 .
  • one or more power systems 230 may be controlled simultaneously.
  • the UAV 200 (UAV) may have multiple horizontally oriented power systems 230 to track the lift and/or push of the target.
  • the horizontally oriented power system 230 may be actuated to provide the ability of the drone 200 (drones) to take off vertically, land vertically, and hover.
  • the UAV 200 may further include a target detection device (not shown in FIG. 2 ), and the target detection device is used to obtain the current frame sensing data collected by the sensor 220 and the target object collected by the sensor 220 .
  • the first candidate area in the sensing data of the previous frame it is also used to determine the second candidate area of the target object in the sensing data of the current frame according to the first candidate area; it is also used to determine the target object according to the second candidate area target detection information.
  • the target detection information includes the category, three-dimensional position coordinates, size, and confidence of the category of the target object.
  • the target detection method provided by the embodiments of the present application will be introduced in detail with reference to the scene in FIG. 1 or FIG. 2 .
  • the scene in FIG. 1 or FIG. 2 is only used to explain the target detection method provided by the embodiment of the present application, but does not constitute a limitation on the application scene of the target detection method provided by the embodiment of the present application.
  • FIG. 3 is a schematic flowchart of steps of a target detection method provided by an embodiment of the present application.
  • the target detection method can be applied to a movable platform to provide stability and accuracy of target detection results.
  • the target detection method includes steps S101 to S103.
  • Step S101 acquiring the sensing data of the current frame and the first candidate region of the target object in the sensing data of the previous frame.
  • the movable platform includes sensors including vision sensors and radar devices, and the sensory data includes image data and/or point cloud data.
  • the sensing data collected by the vision sensor is image data or point cloud data
  • the sensing data collected by the radar device is point cloud data.
  • the collection time of the current frame of sensing data and the previous frame of sensing data differs by a preset time, and the preset time may be set based on the actual situation, which is not specifically limited in this embodiment of the present application.
  • the cache area stores the first candidate area of the target object determined by the movable platform in the previous frame in the sensing data of the previous frame. Therefore, in the current frame, the movable platform can obtain the target from the cache area.
  • the last frame of sensory data collected by the sensor is stored in the cache area. Therefore, in the current frame, the movable platform can perform target detection on the last frame of sensory data to determine whether the target object is in the previous frame.
  • the last frame of sensing data is input into a preset second target detection model to obtain the target detection information of the target object in the last frame and the first candidate region of the target object in the last frame of sensing data.
  • the second target detection model is a pre-trained neural network model
  • the second target detection model is a global-based target detection model
  • the neural network model may include a convolutional neural network model CNN, a recurrent neural network model RNN, and a deep volume
  • the cumulative neural network model DCNN is not specifically limited in this embodiment of the present application.
  • the first candidate area may refer to an area where the target may exist.
  • the number of first candidate regions in a picture may be more than the number of target objects, which is mainly due to different judgment thresholds.
  • target objects with a confidence level greater than 0.7 can be used. It is considered to be the correct detection result, and the target greater than 0.1 is considered to be the correct first candidate region.
  • Step S102 according to the first candidate region, determine a second candidate region of the target object in the sensing data of the current frame.
  • the position of the target object changes little in a short period of time.
  • the first candidate region in the data is directly determined as the second candidate region of the target object in the sensing data of the current frame.
  • the first target detection information corresponding to the first candidate area is obtained; according to the first target detection information and the preset timing prediction algorithm, the second target detection information of the target object in the current frame is predicted; according to the predicted target object The second target detection information of the target object is determined, and the second candidate area of the target object in the sensing data of the current frame is determined.
  • the target objects include cars, pedestrians, traffic lights, lane lines, etc.
  • the timing prediction algorithm the second candidate region of the first candidate region in the sensor data of the current frame can be predicted after a short period of time, thus considering the time sequence information between the sensor data of the current frame and the sensor data of the previous frame, The stability and accuracy of target detection results are greatly improved.
  • the preset timing prediction algorithm may include a mean shift algorithm, a Kalman filter algorithm, a particle filter algorithm, an algorithm for modeling moving objects, etc.
  • the first target detection information may include the category, position coordinates, The confidence of the length, width, height, and category, and the positional reliability of the first candidate area.
  • the second target detection information may include the category, position coordinates, length, width, height, and category confidence of the predicted target object in the current frame. degree, the location reliability of the second candidate region.
  • the method of determining the second candidate area of the target object in the current frame sensing data may be: obtaining the target object in the world coordinate system from the second target detection information. According to the coordinate system conversion relationship between the world coordinate system and the sensor coordinates, the first position coordinates are converted into the second position coordinates of the target object in the sensor coordinate system; the target object is determined according to the second position coordinates. At least one second candidate area in the sensing data of the current frame, that is, a rectangular area with a preset size is formed with the second position coordinate as the center point, and at least one second candidate area is obtained.
  • the preset size may be set based on an actual situation, which is not specifically limited in this embodiment of the present application.
  • the pixel point corresponding to the second position coordinate in the image data 10 is the pixel point 11
  • the pixel point 11 is the center point to form a rectangular area 12 and a rectangular area. 13 and rectangular area 14, therefore, rectangular area 12, rectangular area 13 and rectangular area 14 are the second candidate areas of the target object in the current frame sensing data, the size of rectangular area 12 is smaller than the size of rectangular area 13, rectangular area 13 is smaller than the size of the rectangular area 14 .
  • first target detection information corresponding to each first candidate region is acquired; and according to each first target detection information, multiple first candidate regions are filtered to obtain at least one second candidate region.
  • the first target detection information includes the category, position coordinates, length, width, height, and category confidence of the target object in the previous frame, the location reliability of the first candidate area, and the height of the target object in the second candidate area. Less than or equal to the preset height, and/or, the position coordinates of the target object in the second candidate area are located in the preset position coordinate range, and the preset height and the preset position coordinate range can be set based on the actual situation. This is not specifically limited. Therefore, unreasonable areas are excluded. The unreasonable here mainly refers to object constraints.
  • the vehicle appears on the road by default, and it is impossible to be in the sky. Therefore, the preset position coordinates of the target object The range does not appear in the high sky; for another example, when the target object is a pedestrian, the aspect ratio of the target object is constrained, so the preset height of the pedestrian is 3 meters. Based on the judgment of the above geometric information, the quality of the candidate region is improved, and the robustness of the algorithm can be further improved.
  • Step S103 Determine target detection information of the target object according to the second candidate region.
  • the target detection information includes the category, position coordinates, length, width, height, and category confidence of the target object in the current frame.
  • the travel of the movable platform itself is planned based on the target detection information, and the planning includes at least one of the following: keeping a constant distance from the target object, stopping, and detouring.
  • the second candidate region is input into the preset first target detection model to obtain the target detection information of the target object and the target candidate region of the target object in the sensing data of the current frame.
  • the target candidate area may be one or multiple
  • the first target detection model is a pre-trained neural network model
  • the first target detection model is a local area-based target detection model
  • the training process may be: Acquire a plurality of first training sample data, wherein the first training sample data includes candidate regions of the target object in the sensor data, marked target detection information, and marked candidate regions;
  • the neural network model is iteratively trained until the first neural network model after the iterative training converges, and a first target detection model is obtained.
  • the neural network model may include a convolutional neural network model CNN, a cyclic neural network model RNN, and a deep convolutional neural network model DCNN, which are not specifically limited in this embodiment of the present application.
  • the second candidate region is input into the preset first target detection model to obtain multiple candidate regions of the target object in the sensor data of the current frame and the category confidence of the target object in each candidate region; based on The candidate region whose category confidence is greater than or equal to the first preset confidence level determines the target detection information of the target object, and the candidate region whose category confidence level is greater than or equal to the second preset confidence level is determined as the target object transmitted in the current frame. target candidate regions in the sensory data.
  • the first preset reliability is greater than the second preset reliability, and the first preset reliability and the second preset reliability may be set based on the actual situation, which is not specifically limited in this embodiment of the present application. For example, the first preset reliability is 0.8 and the second preset reliability is 0.2. For another example, the first preset reliability is 0.7 and the second preset reliability is 0.1.
  • the candidate regions of the target object in the current frame sensor data 20 include candidate region 21 , candidate region 22 , candidate region 23 and candidate region 24 , and candidate region 21 , candidate region 22 , candidate region 23 and candidate region 24
  • the category confidences of the region 24 are 0.75, 0.25, 0.8 and 0.1 respectively, and the first preset confidence is 0.8 and the second preset confidence is 0.2. Therefore, based on the candidate region 23 to determine the target detection information of the target object, set the The candidate area 21 , the candidate area 22 , and the candidate area 23 are determined as target candidate areas of the target object in the sensing data of the current frame.
  • the location reliability of the target object in the candidate area with the category confidence greater than or equal to the second preset reliability obtains the location reliability of the target object in the candidate area with the category confidence greater than or equal to the second preset reliability; according to the category confidence greater than or equal to the target in the candidate area with the second preset reliability
  • the location reliability of the object determine the sampling distance of the candidate region whose category confidence is greater than or equal to the second preset reliability; according to the sampling distance of the candidate region whose category confidence is greater than or equal to the second preset reliability and the category
  • the candidate region whose confidence level is greater than or equal to the second preset confidence level is determined as the target candidate region of the target object in the sensing data of the current frame.
  • sampling candidate area and the candidate area of the candidate area 22 can be determined 23 sampling candidate regions, and finally determine candidate region 21, candidate region 22, candidate region 23, sampling candidate region of candidate region 21, sampling candidate region of candidate region 22, and sampling candidate region of candidate region 23 as the target object in the current frame.
  • Target candidate regions in sensory data.
  • the target candidate region of the target object in the current frame of sensory data is displayed. By displaying the target candidate area of the target object in the sensing data of the current frame, it is convenient for users to read and debug.
  • the target candidate area of the target object in the current frame of sensory data is stored in the buffer area. By storing the target candidate area in the cache area, it is convenient to subsequently determine the candidate area of the target object in the next frame of sensory data based on the target candidate area in the cache.
  • step S103 may include: sub-steps S1031 to S1032.
  • Sub-step S1031 According to the preset installation information of the sensor, determine the third candidate area of the target object in the current frame sensing data;
  • Sub-step S1032 Determine target detection information of the target object according to the second candidate area and the third candidate area.
  • the third candidate area of the target object in the sensing data of the current frame can be determined, and then the second candidate area and the third candidate area can determine the target detection information of the target object, thus combining the inspection results of the historical frame
  • the current frame with the prior information of target detection and the installation position, angle and imaging method of the sensor to provide the current frame with the prior information of target detection, and at the same time use the target detection model based on the local area with low complexity to reduce the overall calculation.
  • it can also greatly improve the stability and accuracy of target detection results.
  • the preset position coordinates of the target object in the coordinate system of the sensor are determined; according to the preset position coordinates, the first position of the target object in the current frame sensing data is determined.
  • Three candidate regions Therefore, according to the installation position, angle and imaging method of the sensor, several typical key areas are obtained. For example, the lower left corner of the image captures the position where overtaking may occur in the left lane of the vehicle. This area can be sampled to obtain a Series of candidate regions.
  • the approximate area of the target object can be initially determined, and the candidate area can be obtained by sampling, which improves the quality of the candidate area and further improves the robustness of the algorithm.
  • the approximate area of the target object at the edge of the image can be preliminarily determined, and then the candidate area can be obtained by sampling the approximate area, which can improve the quality of the candidate area, thereby improving the rationality of the target detection result. Robustness of the algorithm.
  • the sum and/or difference between the preset position coordinates and each of the preset position coordinate gains in the multiple preset position coordinate gains are determined to obtain multiple candidate position coordinates; according to the multiple candidate position coordinates and Preset position coordinates to determine multiple third candidate regions of the target object in the current frame of sensing data.
  • the preset installation information is determined according to the installation position of the sensor, and the multiple preset position coordinate gains may be set based on the actual situation, which is not specifically limited in this embodiment of the present application.
  • the pixel point corresponding to the preset position coordinate of the target object in the coordinate system of the sensor is the pixel point 31, and the preset position coordinate and each preset position coordinate in the multiple preset position coordinate gains are obtained.
  • the sum and/or difference between the gains can be obtained as pixel point 32, pixel point 33, and pixel point 34
  • the rectangular area 35 can be determined by the pixel point 31
  • the rectangular area 36 can be determined by the pixel point 32
  • the rectangular area 36 can be determined by the pixel point 33.
  • the rectangular area 37 is determined
  • the rectangular area 38 can be determined by the pixel points 34. Therefore, the rectangular area 35, the rectangular area 36, the rectangular area 37 and the rectangular area 38 are determined as the third candidate area of the target object in the current frame sensing data.
  • the second candidate region and the third candidate region are input into the first target detection model to obtain the target detection information of the target object and the target candidate region of the target object in the current frame of sensory data.
  • the first target detection model is a pre-trained neural network model, and the training process may be as follows: acquiring a plurality of first training sample data, wherein the first training sample data includes candidate regions of the target object in the sensor data , the marked target detection information and the marked candidate area; the first neural network model is iteratively trained according to the plurality of first training sample data, until the first neural network model after the iterative training converges, and the first target detection model is obtained.
  • the target detection information of the target object is determined according to the current frame sensing data and the second candidate region.
  • the second candidate area is input into the preset first target detection model to obtain the third target detection information and the fourth candidate area of the target object in the current frame sensing data;
  • the current frame sensing data is input into the preset obtain the fourth target detection information and the fifth candidate region of the target object in the sensing data of the current frame; according to the third target detection information and the fourth target detection information, determine the target detection of the target object information; according to the fourth candidate area and the fifth candidate area, determine the target candidate area of the target object in the sensing data of the current frame.
  • the target detection information of the target object can be determined more accurately and stably through the current frame sensing data and the second candidate region, which can greatly improve the stability and accuracy of the target detection result.
  • the second target detection model is a pre-trained neural network model, and the second target detection model is a global-based target detection model.
  • the training process may be: acquiring multiple second training sample data, wherein the second training sample The data includes sensor data, marked target detection information and marked candidate regions; the second neural network model is iteratively trained according to multiple second training sample data, until the second neural network model after the iterative training converges, and the second neural network model is obtained.
  • the neural network model may include a convolutional neural network model CNN, a recurrent neural network model RNN and the deep convolutional neural network model DCNN, which is not specifically limited in this embodiment of the present application.
  • the method of determining the target detection information of the target object may be: determining the matching degree between the third target detection information and the fourth target detection information; If the matching degree is greater than or equal to the preset matching degree, the third target detection information and the fourth target detection information are fused to obtain the target detection information of the target object; if the matching degree is less than the preset matching degree, the fourth target detection information is fused
  • the detection information is determined as target detection information of the target object.
  • the degree of matching between the third target detection information and the fourth target detection information includes a category matching degree, a position matching degree and a size matching degree, and the category matching degree is based on the category of the target object in the third target detection information and the first matching degree.
  • the category of the target object in the four target detection information is determined, the position matching degree is determined according to the position coordinates of the target object in the third target detection information and the position coordinates of the target object in the fourth target detection information, and the size matches
  • the degree is determined according to the length, width and height of the target object in the third target detection information and the length, width and height of the target object in the fourth target detection information.
  • the method of fusing the third target detection information and the fourth target detection information to obtain the target detection information of the target object may be: determining the first product of the third target detection information and the first preset coefficient, and determine the second product of the fourth target detection information and the second preset coefficient; determine the sum of the first product and the second product, and determine the sum of the first product and the second product as the target detection information of the target object.
  • the sum of the first preset coefficient and the second preset coefficient is equal to 1, and the first preset coefficient is smaller than the second preset coefficient.
  • the first preset coefficient and the second preset coefficient can be set based on the actual situation. This is not specifically limited in the application examples. For example, the first preset coefficient is 0.4, and the second preset coefficient is 0.6. In another example, the first preset coefficient is 0.45 and the second preset coefficient is 0.55.
  • the similarity between each fourth candidate region and each fifth candidate region is determined; the target candidate region pair is determined from the plurality of fourth candidate regions and the plurality of fifth candidate regions according to the similarity , the target candidate region pair includes a fourth candidate region and a fifth candidate region; the fourth candidate region and/or the fifth candidate region in the target candidate region pair is determined as the target candidate region.
  • the similarity between the fourth candidate region and the fifth candidate region in the pair of target candidate regions is greater than the preset similarity, and the preset similarity may be set based on the actual situation, which is not specifically limited in this embodiment of the present application.
  • the target detection information of the target object is determined according to the current frame sensing data and the second candidate area, that is, the current frame sensing data and the second candidate area.
  • the second target detection model is input to the region to obtain the target detection information of the target object; if the current frame sensor data is not the key frame sensor data, the target detection information of the target object is determined according to the second candidate region, that is, the second candidate region.
  • the first target detection model is input to obtain target detection information of the target object.
  • the first computing resources required for running the first target detection model are smaller than the second computing resources required for running the second target detection model.
  • the accuracy of the target detection result can be ensured by using the second target detection model to determine the target detection result when the sensing data of the current frame is the sensing data of the key frame, and when the sensing data of the current frame is not the sensing data of the key frame , using the first target detection model to determine the target detection result can reduce the consumption of computing resources while ensuring the accuracy of the target detection result.
  • the target detection information of the target object is determined according to the current frame sensing data, the second candidate area and the third candidate area; If the data is not the key frame sensing data, the target detection information of the target object is determined according to the second candidate area and the third candidate area.
  • the target detection result of the previous frame provides a priori information for the target detection of the current frame, and the target detection is performed in combination with the current frame sensor data, which can improve the accuracy of the target detection result and the robustness of the algorithm.
  • the frame number of the current frame sensing data, the target confidence level and/or the remaining computing resources of the target detection information of the target object in the previous frame are obtained; according to the frame number, target confidence level of the current frame sensing data and/or remaining computing resources to determine whether the current frame sensing data is key frame sensing data.
  • the target confidence is determined according to the category confidence and/or location confidence of the target object in each first candidate region.
  • the frame number of the current frame sensing data is an integer multiple of the preset frame number, the target confidence level is less than the preset confidence level, and/or the remaining computing resources are greater than the preset computing resources, it is determined that the current frame sensing data
  • the data is the key frame sensing data; if the frame number of the current frame sensing data is not an integer multiple of the preset frame number, the target confidence level is greater than or equal to the preset confidence level, or the remaining computing resources are less than or equal to the preset computing resources, then It is determined that the sensing data of the current frame is not the sensing data of the key frame.
  • the preset frame number, the preset reliability, and the preset computing resource may be set based on actual conditions, which are not specifically limited in this embodiment of the present application. For example, the preset frame number is 50, and the preset reliability is 0.8.
  • the current frame sensing data is the key frame sensing data is determined by the frame number of the current frame sensing data.
  • the current frame sensing data can be determined as the key frame sensing data at intervals.
  • the global-based second target detection model is used to detect the current frame sensor data, which can ensure the stability and accuracy of the target detection results in terms of time series, and can also detect the first frame in the sensor data time series.
  • the sensing data is regarded as the key frame sensing data, and each subsequent frame of sensing data is regarded as the non-key frame sensing data. Therefore, in a period of time when the current frame sensing data is not the key frame sensing data, using the first target detection model based on the local area to perform target detection on the candidate area can ensure the accuracy of the target detection result and reduce the consumption of computing resources.
  • the target confidence of the target detection information of the target object in the previous frame is less than the preset confidence, that is, when the confidence of the target detection information of the previous frame is relatively low, the movable platform based on the target detection information with lower confidence Planning your own driving is prone to safety accidents and cannot guarantee the safety of the movable platform. Therefore, when the target confidence level of the target detection information of the target object in the previous frame is less than the preset confidence level, the current frame sensing data is determined as the key. frame sensing data, so that the global-based second target detection model is used to perform target detection on the current frame sensing data, which can improve the confidence of the target detection information, so that the mobile platform can plan itself based on accurate target detection information. to avoid safety accidents and improve the safety of movable platforms.
  • the current frame sensing data is determined as the key frame sensing data, so that the second global-based target detection model is used to perform target detection on the current frame sensing data , which can improve the accuracy of target detection information, and when the remaining computing resources of the movable platform are less than or equal to the preset computing resources, the current frame sensing data is determined as non-key frame sensing data, so that the local area-based A target detection model is used to perform target detection on a candidate region, which can reduce the consumption of computing resources while ensuring the accuracy of the target detection result.
  • the method of determining the target confidence of the target detection information of the target object in the previous frame may be: obtaining the category confidence and/or location confidence of the target object in each first candidate area; A category confidence level and/or a location confidence level are determined, and the target confidence level of the target detection information of the target object in the previous frame is determined.
  • each category confidence an average of the category confidences is determined, and the average of the category confidences is determined as the target confidence of the target detection information of the target object in the previous frame.
  • an average value of the fixed position reliability is determined, and the average value of the fixed position reliability is determined as the target confidence degree of the target detection information of the target object in the previous frame.
  • each category confidence determine the average of the category confidence, and determine the average of the category confidence as the first confidence, and at the same time according to each location reliability, determine the average of the location reliability, and use The average value of the location reliability is determined as the second confidence degree, and then the average value of the first confidence degree and the second confidence degree is determined, and the average value of the first confidence degree and the second confidence degree is determined as the target object in the previous frame.
  • the target confidence of the target detection information determine the average of the category confidence, and determine the average of the category confidence as the first confidence, and at the same time according to each location reliability, determine the average of the location reliability, and use The average value of the location reliability is determined as the second confidence degree, and then the average value of the first confidence degree and the second confidence degree is determined, and the average value of the first confidence degree and the second confidence degree is determined as the target object in the previous frame.
  • the target confidence of the target detection information determines the average value of the category confidence, and determine the average of the category confidence as the first confidence, and at the same time according to each
  • the current frame of sensory data and the first candidate region of the target object in the previous frame of sensory data are obtained; according to the first candidate region, the second candidate region of the target object in the current frame of sensory data is determined. ; Obtain the target detection information of the target object in the previous frame, and predict the target detection information of the target object in the current frame based on the preset timing prediction algorithm and the target detection information of the target object in the previous frame; The target detection information of the frame is used to determine the fourth candidate area of the target object in the sensing data of the current frame; the target detection information of the target object in the current frame is determined according to the second candidate area and the fourth candidate area.
  • the candidate area of the target object in the current frame can be predicted, the quality of the candidate area can be improved, the accuracy of the target detection result and the robustness of the algorithm can be improved.
  • the target detection information of the target object in the current frame may also be determined according to the second candidate area, the third candidate area, and the fourth candidate area.
  • the target detection information of the target object in the current frame may also be determined according to the second candidate area, the fourth candidate area and the current frame sensing data.
  • the target detection information of the target object in the current frame may also be determined according to the second candidate area, the third candidate area, the fourth candidate area and the current frame sensing data. This embodiment of the present application does not specifically limit this.
  • the current frame sensing data and the first candidate region of the target object in the previous frame of the sensing data are obtained, and then according to the first candidate region, it is determined that the target object is in the current frame sensing data.
  • the second candidate area of the The timing information between the frame sensing data and the previous frame sensing data greatly improves the stability and accuracy of target detection results.
  • FIG. 9 is a schematic structural block diagram of a target detection apparatus provided by an embodiment of the present application.
  • the target detection apparatus 200 includes a processor 201 and a memory 202, and the processor 201 and the memory 202 are connected through a bus 203, such as an I2C (Inter-integrated Circuit) bus.
  • a bus 203 such as an I2C (Inter-integrated Circuit) bus.
  • the processor 201 may be a micro-controller unit (Micro-controller Unit, MCU), a central processing unit (Central Processing Unit, CPU), or a digital signal processor (Digital Signal Processor, DSP) or the like.
  • MCU Micro-controller Unit
  • CPU Central Processing Unit
  • DSP Digital Signal Processor
  • the memory 202 may be a Flash chip, a read-only memory (ROM, Read-Only Memory) magnetic disk, an optical disk, a U disk, or a removable hard disk, or the like.
  • ROM Read-Only Memory
  • the memory 202 may be a Flash chip, a read-only memory (ROM, Read-Only Memory) magnetic disk, an optical disk, a U disk, or a removable hard disk, or the like.
  • the processor 201 is used for running the computer program stored in the memory 202, and implements the following steps when executing the computer program:
  • the first candidate area determine the second candidate area of the target object in the current frame sensor data
  • target detection information of the target object is determined.
  • the processor determines the second candidate region of the target object in the sensing data of the current frame according to the first candidate region, the processor is configured to:
  • a plurality of the first candidate regions are filtered to obtain at least one of the second candidate regions.
  • the first target detection information includes the height and position coordinates of the target object, the height of the target object in the second candidate area is less than or equal to a preset height, and/or , the position coordinates of the target object in the second candidate area are within a preset position coordinate range.
  • the processor when the processor determines the second candidate region of the target object in the sensing data of the current frame according to the first candidate region, the processor is configured to:
  • the first target detection information and the preset timing prediction algorithm predict the second target detection information of the target object in the current frame
  • a second candidate region of the target object in the current frame sensing data is determined.
  • the processor when the processor determines the target detection information of the target object according to the second candidate region, the processor is configured to:
  • the processor is further configured to implement the following steps:
  • the first training sample data includes candidate regions of the target object in the sensor data, marked target detection information, and marked candidate regions;
  • the first neural network model is iteratively trained according to the plurality of first training sample data, until the first neural network model after the iterative training converges, and the first target detection model is obtained.
  • the processor is further configured to implement the following steps:
  • the target candidate area is displayed.
  • the processor when the processor determines the target detection information of the target object according to the second candidate region, the processor is configured to:
  • the preset installation information of the sensor determine the third candidate area of the target object in the current frame sensing data
  • target detection information of the target object is determined.
  • the processor when the processor determines the third candidate area of the target object in the current frame sensing data according to the preset installation information of the sensor, the processor is configured to:
  • the preset installation information of the sensor determine the preset position coordinates of the target object in the coordinate system of the sensor
  • a third candidate region of the target object in the current frame sensing data is determined.
  • the processor when the processor determines the third candidate region of the target object in the current frame of sensing data according to the preset position coordinates, the processor is configured to:
  • a plurality of third candidate regions of the target object in the sensing data of the current frame are determined.
  • the processor when the processor determines the target detection information of the target object according to the second candidate region and the third candidate region, the processor is configured to:
  • the second candidate area and the third candidate area are input into the first target detection model to obtain the target detection information of the target object and the target candidate area of the target object in the current frame sensing data.
  • the processor when the processor determines the target detection information of the target object according to the second candidate region, the processor is configured to:
  • target detection information of the target object is determined.
  • the processor when the processor determines the target detection information of the target object according to the current frame sensing data and the second candidate region, the processor is configured to:
  • a target candidate area of the target object in the current frame sensing data is determined.
  • the processor when the processor determines the target detection information of the target object according to the third target detection information and the fourth target detection information, the processor is configured to:
  • the third target detection information and the fourth target detection information are fused to obtain target detection information of the target object.
  • the processor is further configured to implement the following steps:
  • the fourth target detection information is determined as the target detection information of the target object.
  • the fourth candidate area and the fifth candidate area are both multiple, and the processor determines the target object according to the fourth candidate area and the fifth candidate area.
  • the target candidate area in the current frame sensing data it is used to realize:
  • a target candidate region pair is determined from a plurality of the fourth candidate regions and a plurality of the fifth candidate regions according to the similarity, and the target candidate region pair includes one of the fourth candidate regions and one of the fifth candidate regions candidate area;
  • the fourth candidate region and/or the fifth candidate region in the pair of target candidate regions is determined as the target candidate region.
  • the similarity between the fourth candidate region and the fifth candidate region in the pair of target candidate regions is greater than a preset similarity.
  • the processor is further configured to implement the following steps:
  • the second training sample data includes sensor data, marked target detection information and marked candidate regions;
  • the second neural network model is iteratively trained according to the plurality of second training sample data, until the iteratively trained second neural network model converges, and the second target detection model is obtained.
  • the first computing resources required to run the first object detection model are smaller than the second computing resources required to run the second object detection model.
  • the processor is further configured to implement the following steps:
  • target detection information of the target object is determined according to the current frame sensing data and the second candidate region.
  • the processor is further configured to implement the following steps:
  • the target confidence level and/or the remaining computing resources it is determined whether the current frame sensing data is key frame sensing data.
  • the processor determines whether the current frame sensing data is a key frame sensing data according to the frame number of the current frame sensing data, the target confidence level and/or the remaining computing resources. data, used to achieve:
  • the sensor data is key frame sensor data.
  • the processor is further configured to implement the following steps:
  • the target confidence of the target detection information of the target object in the previous frame is determined according to each of the category confidence and/or location confidence.
  • the sensors include vision sensors and radar devices.
  • FIG. 10 is a schematic structural block diagram of a movable platform provided by an embodiment of the present application.
  • the movable platform 300 includes a platform body 310 , a power system 320 , a sensor 330 and a target detection device 340 .
  • the power system 320 , the sensor 330 and the target detection device 340 are provided on the platform body 310 , and the power system 320 is used for
  • the movable platform 300 is provided with moving power, the sensor 330 is used for collecting sensing data, and the target detection device 340 is used for determining the target detection information of the target object and also for controlling the movable platform 300 .
  • the movable platform 300 includes unmanned aerial vehicles, robots, unmanned boats, unmanned vehicles, and the like.
  • Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, the computer program includes program instructions, and the processor executes the program instructions, so as to realize the provision of the above embodiments.
  • the steps of the object detection method are described in detail below.
  • the computer-readable storage medium may be an internal storage unit of the removable platform described in any of the foregoing embodiments, such as a hard disk or a memory of the removable platform.
  • the computer-readable storage medium can also be an external storage device of the removable platform, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital) equipped on the removable platform , SD) card, flash memory card (Flash Card), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Optical Radar Systems And Details Thereof (AREA)

Abstract

一种目标检测方法、装置、可移动平台及计算机可读存储介质，其中该方法包括：获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域(S101)；根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域(S102)；根据所述第二候选区域，确定所述目标对象的目标检测信息(S103)。该方法能够提高目标检测结果的稳定性和准确性。

Description

目标检测方法、装置、可移动平台及计算机可读存储介质 技术领域
本申请涉及目标检测技术领域，尤其涉及一种目标检测方法、装置、可移动平台及计算机可读存储介质。
背景技术
对于可移动平台的控制，需要获取可移动平台所处环境的目标对象的目标检测结果，通过目标对象的目标检测结果可以控制可移动平台避开可移动平台移动方向上的障碍物，保证可移动平台的安全。目前，可移动平台主要是控制传感器连续的采集传感数据，并通过目标检测算法对每一帧传感数据进行单独处理，得到每一帧传感数据的处理结果，然后再对每一帧传感数据的处理结果进行融合处理，从而可以得到目标对象的目标检测结果。然而，对每一帧传感数据进行单独处理，无法保证目标检测结果的稳定性和准确性，因此，目标检测结果的稳定性和准确性还有待提高。
发明内容
基于此，本申请实施例提供了一种目标检测方法、装置、可移动平台及计算机可读存储介质，旨在提高目标检测结果的稳定性和准确性。
第一方面，本申请实施例提供了一种目标检测方法，包括：
获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；
根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域；
根据所述第二候选区域，确定所述目标对象的目标检测信息。
第二方面，本申请实施例还提供了一种目标检测装置，所述目标检测装置包括存储器和处理器；
所述存储器用于存储计算机程序；
所述处理器，用于执行所述计算机程序并在执行所述计算机程序时，实现如下步骤：
获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；
根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域；
根据所述第二候选区域，确定所述目标对象的目标检测信息。
第三方面，本申请实施例还提供了一种可移动平台，包括：
平台本体；
动力系统，设于所述平台本体上，用于为所述可移动平台提供移动动力；
传感器，设于所述平台本体上，用于采集传感数据；
如上所述目标检测装置，设于所述平台本体内，用于确定目标对象的目标检测信息以及还用于控制所述可移动平台。
第四方面，本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时使所述处理器在实现如上所述的目标检测方法的步骤。
本申请实施例提供了一种目标检测方法、装置、可移动平台及计算机可读存储介质，通过获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域，然后根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域，最后根据第二候选区域，确定目标对象的目标检测信息，由于第二候选区域是基于目标对象在上一帧传感数据中的第一候选区域确定的，因此考虑了当前帧传感数据与上一帧传感数据之间的时序信息，极大的提高了目标检测结果的稳定性和准确性。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，并不能限制本申请。
附图说明
为了更清楚地说明本申请实施例技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是实施本申请实施例提供的目标检测方法的一场景示意图；
图2是实施本申请实施例提供的目标检测方法的另一场景示意图；
图3是本申请实施例提供的一种目标检测方法的步骤示意流程图；
图4是本申请实施例中确定第二候选区域的一场景示意图；
图5是本申请实施例中确定目标候选区域的一场景示意图；
图6是本申请实施例中确定目标候选区域的另一场景示意图；
图7是图3中的目标检测方法的子步骤示意流程图；
图8是本申请实施例中确定第三候选区域的一场景示意图；
图9是本申请实施例提供的一种目标检测装置的结构示意性框图；
图10是本申请实施例提供的一种可移动平台的结构示意性框图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
附图中所示的流程图仅是示例说明，不是必须包括所有的内容和操作/步骤，也不是必须按所描述的顺序执行。例如，有的操作/步骤还可以分解、组合或部分合并，因此实际执行的顺序有可能根据实际情况改变。
下面结合附图，对本申请的一些实施方式作详细说明。在不冲突的情况下，下述的实施例及实施例中的特征可以相互组合。
对于可移动平台的控制，需要获取可移动平台所处环境的目标对象的目标检测结果，通过目标对象的目标检测结果可以控制可移动平台避开可移动平台移动方向上的障碍物，保证可移动平台的安全。目前，可移动平台主要是控制传感器连续的采集传感数据，并通过目标检测算法对每一帧传感数据进行单独处理，得到每一帧传感数据的处理结果，然后再对每一帧传感数据的处理结果进行融合处理，从而可以得到目标对象的目标检测结果。然而，对每一帧传感数据进行单独处理，并没有考虑传感数据之间的时序信息，因此，目标检测结果的稳定性和准确性还有待提高。
为解决上述问题，本申请实施例提供了一种目标检测方法、装置、可移动平台及计算机可读存储介质。通过获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域，然后根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域，最后根据第二候选区域，确定目标对象的目标检测信息，由于第二候选区域是基于目标对象在上一帧传感数据中的第一候选区域确定的，因此考虑了当前帧传感数据与上一帧传感数据之间的时序信息，极大的提高了目标检测结果的稳定性和准确性。
在一实施例中，该目标检测方法可以应用于可移动平台，可移动平台包括无人机、机器人、无人船和无人驾驶汽车等。请参阅图1，图1是实施本申请实施例提供的目标检测方法的一场景示意图。如图1所示，无人驾驶汽车100 包括汽车本体110、设于汽车本体110上的传感器120和设于汽车本体110上的动力系统130，传感器120用于采集传感数据，动力系统130用于为无人驾驶汽车100提供移动动力。
其中，传感器120包括视觉传感器和雷达装置，视觉传感器可以是单目视觉传感器，也可以是双目视觉传感器，雷达装置可以包括激光雷达、毫米波雷达。可选的，无人驾驶汽车100可以包括一个或多个雷达装置。以激光雷达为例，激光雷达可以通过发射激光束探测某个环境中物体的位置、速度等信息，从而获得激光点云。激光雷达可以向包括目标物的环境发射探测信号，然后接受从目标物反射回来的反射信号，根据反射的探测信号、接收到的反射信号，并根据发送和接收的间隔时间等数据参数，获得激光点云。激光点云可以包括N个点，每个点可以包括x,y,z坐标和intensity(反射率)等参数值。
在一实施例中，无人驾驶汽车100还可以包括目标检测装置(图1中未示出)，目标检测装置用于获取传感器120采集到的当前帧传感数据和目标对象在传感器120采集到的上一帧传感数据中的第一候选区域；还用于根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域；还用于根据第二候选区域，确定目标对象的目标检测信息。其中，目标检测信息包括目标对象的类别、三维位置坐标、尺寸、该类别的置信度等。
请参阅图2，图2是实施本申请实施例提供的目标检测方法的另一场景示意图。如图2所示，无人机200包括机体210、设于机体210上的传感器220和设于机体210上的动力系统230，传感器220用于采集传感数据，动力系统230用于为无人机200提供飞行动力。其中，传感器220包括视觉传感器和雷达装置，雷达装置可以包括激光雷达、毫米波雷达。可选的，无人机200可以包括一个或多个雷达装置。
其中，水平方向的动力系统230中的一个或者多个可以顺时针方向旋转，而水平方向的动力系统230中的其它一个或者多个可以逆时针方向旋转。例如，顺时针旋转的动力系统230与逆时针旋转的动力系统230的数量一样。每一个水平方向的动力系统230的旋转速率可以独立变化，以实现每个动力系统230导致的提升及/或推动操作，从而调整无人机200的空间方位、速度及/或加速度(如相对于多达三个自由度的旋转及平移)。
在一实施例中，动力系统230能够使无人机200(无人机)垂直地从地面起飞，或者垂直地降落在地面上，而不需要无人机200(无人机)任何水平运动(如不需要在跑道上滑行)。可选的，动力系统230可以允许无人机200(无 人机)在空中预设位置和/或方向盘旋。一个或者多个动力系统230在受到控制时可以独立于其它的动力系统230。可选的，一个或者多个动力系统230可以同时受到控制。例如，无人机200(无人机)可以有多个水平方向的动力系统230，以追踪目标的提升及/或推动。水平方向的动力系统230可以被致动以提供无人机200(无人机)垂直起飞、垂直降落、盘旋的能力。
在一实施例中，无人机200还可以包括目标检测装置(图2中未示出)，目标检测装置用于获取传感器220采集到的当前帧传感数据和目标对象在传感器220采集到的上一帧传感数据中的第一候选区域；还用于根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域；还用于根据第二候选区域，确定目标对象的目标检测信息。其中，目标检测信息包括目标对象的类别、三维位置坐标、尺寸、该类别的置信度等。
以下，将结合图1或图2中的场景对本申请的实施例提供的目标检测方法进行详细介绍。需知，图1或图2中的场景仅用于解释本申请实施例提供的目标检测方法，但并不构成对本申请实施例提供的目标检测方法应用场景的限定。
请参阅图3，图3是本申请实施例提供的一种目标检测方法的步骤示意流程图。该目标检测方法可以应用在可移动平台，用于提供目标检测结果的稳定性和准确性。
具体地，如图3所示，该目标检测方法包括步骤S101至步骤S103。
步骤S101、获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域。
示例性的，可移动平台包括传感器，该传感器包括视觉传感器和雷达装置，传感数据包括图像数据和/或点云数据。例如，若传感器为视觉传感器，则视觉传感器采集到的传感数据为图像数据或点云数据，若传感器为雷达装置，则雷达装置采集到的传感数据为点云数据。当前帧传感数据与上一帧传感数据的采集时间相差预设时间，该预设时间可基于实际情况进行设置，本申请实施例对此不做具体限定。
示例性的，缓存区域中存储有可移动平台在上一帧确定的目标对象在上一帧传感数据中的第一候选区域，因此，在当前帧，可移动平台可以从缓存区域中获取目标对象在上一帧传感数据中的第一候选区域。通过将目标对象在上一帧传感数据中的第一候选区域存储在缓存区域中，便于后续从缓存区域中读取，不需要耗费计算资源再次进行计算，减少计算资源的消耗。
在一实施例中，缓存区域中存储有传感器采集到的上一帧传感数据，因此， 在当前帧，可移动平台可以对上一帧传感数据进行目标检测，以确定目标对象在上一帧传感数据中的第一候选区域。示例性的，将上一帧传感数据输入预设的第二目标检测模型，得到目标对象的在上一帧的目标检测信息和目标对象在上一帧传感数据中的第一候选区域。其中，第二目标检测模型为预先训练好的神经网络模型，第二目标检测模型为基于全局的目标检测模型，该神经网络模型可以包括卷积神经网络模型CNN、循环神经网络模型RNN和深度卷积神经网络模型DCNN，本申请实施例对此不做具体限定。
在一实施例中，第一候选区域可以指可能存在目标的区域。以普通的驾驶场景、传感器为视觉传感器为例，一张图片中第一候选区域的数量可能多于目标对象的数量，这主要由于其判断阈值不同，例如，可以将置信度大于0.7的目标对象认为是正确的检测结果，把大于0.1的目标认为是正确的第一候选区域。
步骤S102、根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域。
在一实施例中，由于当前帧传感数据与上一帧传感数据的间隔时间较短，在短时间内，目标对象的位置变化较小，因此，可以将目标对象在上一帧传感数据中的第一候选区域直接确定为目标对象在当前帧传感数据中的第二候选区域。利用历史帧的检测结果可以为当前帧的检测提供先验信息，极大的提高了目标检测结果的稳定性和准确性。
在一实施例中，获取第一候选区域对应的第一目标检测信息；根据第一目标检测信息和预设时序预测算法，预测目标对象在当前帧的第二目标检测信息；根据预测的目标对象的第二目标检测信息，确定目标对象在当前帧传感数据中的第二候选区域。其中，目标对象包括汽车、行人、交通灯、车道线等。通过时序预测算法可以预测第一候选区域经过一段较短时间后在当前帧传感数据中的第二候选区域，从而考虑了当前帧传感数据与上一帧传感数据之间的时序信息，极大的提高了目标检测结果的稳定性和准确性。
其中，预设时序预测算法可以包括均值漂移算法、Kalman滤波算法、粒子滤波算法、对运动目标建模算法等，其中，第一目标检测信息可以包括目标对象在上一帧的类别、位置坐标、长度、宽度、高度、类别的置信度、第一候选区域的定位置信度，第二目标检测信息可以包括预测得到的目标对象在当前帧的类别、位置坐标、长度、宽度、高度、类别的置信度、第二候选区域的定位置信度。
在一实施例中，根据预测的第二目标检测信息，确定目标对象在当前帧传感数据中的第二候选区域的方式可以为：从第二目标检测信息中获取目标对象在世界坐标系下的第二位置坐标，并按照世界坐标系与传感器坐标之间的坐标系转换关系，将第一位置坐标转换为目标对象在传感器坐标系中的第二位置坐标；根据第二位置坐标确定目标对象在当前帧传感数据中的至少一个第二候选区域，即以第二位置坐标为中心点形成预设尺寸的矩形区域，得到至少一个第二候选区域。其中，预设尺寸可基于实际情况进行设置，本申请实施例对此不做具体限定。
以当前帧传感数据为图像数据为例，如图4所示，第二位置坐标在图像数据10中的对应像素点为像素点11，以像素点11为中心点形成矩形区域12、矩形区域13和矩形区域14，因此，矩形区域12、矩形区域13和矩形区域14为目标对象在当前帧传感数据中的第二候选区域，矩形区域12的尺寸小于矩形区域13的尺寸，矩形区域13的尺寸小于矩形区域14的尺寸。
在一实施例中，获取每个第一候选区域对应的第一目标检测信息；根据每个第一目标检测信息，对多个第一候选区域进行过滤，得到至少一个第二候选区域。其中，第一目标检测信息包括目标对象在上一帧的类别、位置坐标、长度、宽度、高度、类别的置信度、第一候选区域的定位置信度，第二候选区域中的目标对象的高度小于或等于预设高度，和/或，第二候选区域中的目标对象的位置坐标位于预设位置坐标范围，预设高度和预设位置坐标范围可基于实际情况进行设置，本申请实施例对此不做具体限定。从而，排除掉不合理的区域，此处的不合理主要指的是物体约束，例如，对于传感器设置于车辆的场景，车辆默认出现在路面上，不可能在天上，因此目标对象预设位置坐标范围不会出现在高空中；又例如，当目标对象为行人时，目标对象的宽高比存在约束，因此行人的预设高度为3米。基于对上述几何信息的判断，提升了候选区域的质量，可以进一步提升算法的鲁棒性。
步骤S103、根据所述第二候选区域，确定所述目标对象的目标检测信息。
其中，目标检测信息包括目标对象在当前帧的类别、位置坐标、长度、宽度、高度、类别的置信度。通过目标检测信息，基于目标检测信息对可移动平台自身的行驶进行规划，所述规划包括如下至少一种：与目标对象保持恒定距离行进、停止行进、绕道行进。
在一实施例中，将第二候选区域输入预设的第一目标检测模型，得到目标对象的目标检测信息和目标对象在当前帧传感数据中的目标候选区域。其中， 目标候选区域可以为一个，也可以为多个，第一目标检测模型为预先训练好的的神经网络模型，第一目标检测模型为基于局部区域的目标检测模型，其训练过程可以为：获取多个第一训练样本数据，其中，第一训练样本数据包括目标对象在传感数据中的候选区域、标注的目标检测信息和标注的候选区域；根据多个第一训练样本数据对第一神经网络模型进行迭代训练，直到迭代训练后的第一神经网络模型收敛，得到第一目标检测模型。该神经网络模型可以包括卷积神经网络模型CNN、循环神经网络模型RNN和深度卷积神经网络模型DCNN，本申请实施例对此不做具体限定。
示例性的，将第二候选区域输入预设的第一目标检测模型，得到目标对象的在当前帧传感数据中的多个候选区域和每个候选区域中的目标对象的类别置信度；基于该类别置信度大于或等于第一预设置信度的候选区域确定目标对象的目标检测信息，并将该类别置信度大于或等于第二预设置信度的候选区域确定为目标对象在当前帧传感数据中的目标候选区域。其中，第一预设置信度大于第二预设置信度，第一预设置信度和第二预设置信度可基于实际情况进行设置，本申请实施例对此不做具体限定。例如，第一预设置信度为0.8，第二预设置信度为0.2，又例如，第一预设置信度为0.7，第二预设置信度为0.1。
如图5所示，目标对象在当前帧传感数据20中的候选区域包括候选区域21、候选区域22、候选区域23和候选区域24，且候选区域21、候选区域22、候选区域23和候选区域24的类别置信度分别为0.75、0.25、0.8和0.1，且第一预设置信度为0.8，第二预设置信度为0.2，因此，基于候选区域23确定目标对象的目标检测信息，将候选区域21、候选区域22、候选区域23确定为目标对象在当前帧传感数据中的目标候选区域。
示例性的，获取该类别置信度大于或等于第二预设置信度的候选区域中的目标对象的定位置信度；根据该类别置信度大于或等于第二预设置信度的候选区域中的目标对象的定位置信度，确定该类别置信度大于或等于第二预设置信度的候选区域的采样距离；根据该类别置信度大于或等于第二预设置信度的候选区域的采样距离和该类别置信度大于或等于第二预设置信度的候选区域，确定目标对象在当前帧传感数据中的目标候选区域。
例如，如图6所示，候选区域21定位置信度分别为0.8，则候选区域21的采样距离为0.8*100＝20像素，候选区域21中的像素点211、像素点212、像素点213和像素点214的像素坐标分别为(x 1,y 1)、(x 2,y 2)、(x 3,y 3)、(x 4,y 4)，则像素点(x 1+20,y 1)、(x 2+20,y 2)、(x 3+20,y 3)、(x 4+20,y 4)围成的矩形区域、像素点 (x 1-20,y 1)、(x 2-20,y 2)、(x 3-20,y 3)、(x 4-20,y 4)围成的矩形区域、(x 1,y 1+20)、(x 2,y 2+20)、(x 3,y 3+20)、(x 4,y 4+20)围成的矩形区域、(x 1,y 1-20)、(x 2,y 2-20)、(x 3,y 3-20)、(x 4,y 4-20)围成的矩形区域为候选区域21的采样候选区域，按照类似的方式，可以确定候选区域22的采样候选区域和候选区域23的采样候选区域，最后将候选区域21、候选区域22、候选区域23、候选区域21的采样候选区域、候选区域22的采样候选区域和候选区域23的采样候选区域确定为目标对象在当前帧传感数据中的目标候选区域。
在一实施例中，显示目标对象在当前帧传感数据中的目标候选区域。通过显示目标对象在当前帧传感数据中的目标候选区域，便于用户阅览，从而方便用户调试。在一实施例中，将目标对象在当前帧传感数据中的目标候选区域存储至缓存区域中。通过将目标候选区域存储至缓存区域中，便于后续基于缓存中的目标候选区域确定目标对象在下一帧传感数据中的候选区域。
在一实施例中，如图7所示，步骤S103可以包括：子步骤S1031至S1032。
子步骤S1031、根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域；
子步骤S1032、根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息。
通过传感器的预设安装信息能够确定目标对象在当前帧传感数据中的第三候选区域，之后第二候选区域和第三候选区域确定目标对象的目标检测信息，从而结合了历史帧的检查结果给当前帧提供目标检测的先验信息以及传感器的安装位置、角度和成像方式给当前帧提供目标检测的先验信息，同时使用复杂度较低的基于局部区域的目标检测模型，在降低总体计算资源消耗的同时，也能够极大的提高目标检测结果的稳定性和准确性。
在一实施例中，根据该传感器的预设安装信息，确定该目标对象在传感器的坐标系下的预设位置坐标；根据该预设位置坐标，确定目标对象在当前帧传感数据中的第三候选区域。从而根据传感器的安装位置、角度和成像方式，得到典型的几个关键区域，例如图像的左下角拍摄到的是本车左侧车道可能出现超车的位置，则可以将此区域进行采样，得到一系列的候选区域。对于远处的目标对象，可以初步确定目标对象的大致区域，通过采样得到候选区域，提升了候选区域的质量，可以进一步提升算法的鲁棒性。对于从图像边缘进入图像内的目标对象，可以初步确定目标对象在图像边缘的大致区域，再通过对大致区域进行采样得到候选区域，可以提升候选区域的质量，进而提高目标检测结 果的合理性和算法的鲁棒性。
具体的，确定预设位置坐标与多个预设位置坐标增益中的每个预设位置坐标增益之间的和值和/或差值，得到多个候选位置坐标；根据多个候选位置坐标和预设位置坐标，确定目标对象在当前帧传感数据中的多个第三候选区域。其中，预设安装信息是根据传感器的安装位置确定的，多个预设位置坐标增益可基于实际情况进行设置，本申请实施例对此不做具体限定。
如图8所示，该目标对象在传感器的坐标系下的预设位置坐标对应的像素点为像素点31，通过预设位置坐标与多个预设位置坐标增益中的每个预设位置坐标增益之间的和值和/或差值可以得到像素点32、像素点33、像素点34，通过像素点31可以确定矩形区域35，通过像素点32可以确定矩形区域36，通过像素点33可以确定矩形区域37，通过像素点34可以确定矩形区域38，因此，将矩形区域35、矩形区域36、矩形区域37和矩形区域38确定为目标对象在当前帧传感数据中的第三候选区域。
示例性的，将第二候选区域和第三候选区域输入第一目标检测模型，得到目标对象的目标检测信息和目标对象在所述当前帧传感数据中的目标候选区域。其中，第一目标检测模型为预先训练好的的神经网络模型，其训练过程可以为：获取多个第一训练样本数据，其中，第一训练样本数据包括目标对象在传感数据中的候选区域、标注的目标检测信息和标注的候选区域；根据多个第一训练样本数据对第一神经网络模型进行迭代训练，直到迭代训练后的第一神经网络模型收敛，得到第一目标检测模型。
在一实施例中，根据当前帧传感数据和第二候选区域，确定目标对象的目标检测信息。示例性的，将第二候选区域输入预设的第一目标检测模型，得到第三目标检测信息和目标对象在当前帧传感数据中的第四候选区域；将当前帧传感数据输入预设的第二目标检测模型，得到第四目标检测信息和目标对象在所述当前帧传感数据中的第五候选区域；根据第三目标检测信息和第四目标检测信息，确定目标对象的目标检测信息；根据第四候选区域和第五候选区域，确定目标对象在当前帧传感数据中的目标候选区域。通过当前帧传感数据和第二候选区域可以更加准确和稳定的确定目标对象的目标检测信息，能够极大的提高目标检测结果的稳定性和准确性。
其中，第二目标检测模型为预先训练好的神经网络模型，第二目标检测模型为基于全局的目标检测模型，其训练过程可以为：获取多个第二训练样本数据，其中，第二训练样本数据包括传感数据，标注的目标检测信息和标注的候 选区域；根据多个第二训练样本数据对第二神经网络模型进行迭代训练，直到迭代训练后的第二神经网络模型收敛，得到第二目标检测模型，运行第一目标检测模型所需的第一计算资源小于运行第二目标检测模型所需的第二计算资源，该神经网络模型可以包括卷积神经网络模型CNN、循环神经网络模型RNN和深度卷积神经网络模型DCNN，本申请实施例对此不做具体限定。
在一实施例中，根据第三目标检测信息和第四目标检测信息，确定目标对象的目标检测信息的方式可以为：确定第三目标检测信息与第四目标检测信息之间的匹配度；若该匹配度大于或等于预设匹配度，则对第三目标检测信息和第四目标检测信息进行融合，得到目标对象的目标检测信息；若该匹配度小于预设匹配度，则将第四目标检测信息确定为目标对象的目标检测信息。通过在该匹配度较高时对第三目标检测信息和第四目标检测信息进行融合，可以得到更加准确的目标检测信息，提高目标检测结果的稳定性和准确性，而在匹配度较低时，将基于当前帧传感数据确定的第四目标检测信息确定为目标对象的目标检测信息，也可以保证目标检测结果的准确性。
其中，第三目标检测信息与第四目标检测信息之间的匹配度包括类别匹配度、位置匹配度和尺寸匹配度，该类别匹配度是根据第三目标检测信息中的目标对象的类别与第四目标检测信息中的目标对象的类别确定的，该位置匹配度是根据第三目标检测信息中的目标对象的位置坐标与第四目标检测信息中的目标对象的位置坐标确定的，该尺寸匹配度是根据第三目标检测信息中的目标对象的长宽高与第四目标检测信息中的目标对象的长宽高确定的。
在一实施例中，对第三目标检测信息和第四目标检测信息进行融合，得到目标对象的目标检测信息的方式可以为：确定第三目标检测信息与第一预设系数的第一乘积，并确定第四目标检测信息与第二预设系数的第二乘积；确定第一乘积与第二乘积之和，并将第一乘积与第二乘积之和确定为目标对象的目标检测信息。其中，第一预设系数与第二预设系数之和等于1，且第一预设系数小于第二预设系数，第一预设系数与第二预设系数可基于实际情况进行设置，本申请实施例对此不做具体限定。例如，第一预设系数为0.4，第二预设系数为0.6，又例如，第一预设系数为0.45，第二预设系数为0.55。
在一实施例中，确定每个第四候选区域与每个第五候选区域之间的相似度；根据该相似度从多个第四候选区域和多个第五候选区域中确定目标候选区域对，该目标候选区域对包括一个第四候选区域和一个第五候选区域；将目标候选区域对中的第四候选区域和/或第五候选区域确定为目标候选区域。其中，目标候 选区域对中的第四候选区域与第五候选区域的相似度大于预设相似度，预设相似度可基于实际情况进行设置，本申请实施例对此不做具体限定。通过从多个第四候选区域和多个第五候选区域确定相似度较高的候选区域，可以提高目标候选区域的准确性，从而提高目标检测结果的准确性和稳定性。
在一实施例中，若当前帧传感数据为关键帧传感数据，则根据当前帧传感数据和第二候选区域，确定目标对象的目标检测信息，即将当前帧传感数据和第二候选区域输入第二目标检测模型，得到目标对象的目标检测信息；若当前帧传感数据不为关键帧传感数据，则根据第二候选区域，确定目标对象的目标检测信息，即将第二候选区域输入第一目标检测模型，得到目标对象的目标检测信息。其中，运行第一目标检测模型所需的第一计算资源小于运行第二目标检测模型所需的第二计算资源。通过在当前帧传感数据为关键帧传感数据时，使用第二目标检测模型来确定目标检测结果可以保证目标检测结果的准确性，而在当前帧传感数据不为关键帧传感数据时，使用第一目标检测模型来确定目标检测结果，可以在保证目标检测结果的准确性的同时，减少计算资源的消耗。
在一实施例中，若当前帧传感数据为关键帧传感数据，则根据当前帧传感数据、第二候选区域和第三候选区域，确定目标对象的目标检测信息；若当前帧传感数据不为关键帧传感数据，则根据第二候选区域和第三候选区域，确定目标对象的目标检测信息。通过上一帧的目标检测结果给当前帧的目标检测提供先验信息，同时结合当前帧传感数据进行目标检测，可以提高目标检测结果的准确性和算法的鲁棒性。
在一实施例中，获取当前帧传感数据的帧号、目标对象在上一帧的目标检测信息的目标置信度和/或剩余计算资源；根据当前帧传感数据的帧号、目标置信度和/或剩余计算资源，确定当前帧传感数据是否为关键帧传感数据。其中，该目标置信度是根据每个第一候选区域中的目标对象的类别置信度和/或定位置信度确定的。
在一实施例中，若当前帧传感数据的帧号为预设帧号的整数倍、目标置信度小于预设置信度和/或剩余计算资源大于预设计算资源，则确定当前帧传感数据为关键帧传感数据；若当前帧传感数据的帧号不为预设帧号的整数倍、目标置信度大于或等于预设置信度或剩余计算资源小于或等于预设计算资源，则确定当前帧传感数据不为关键帧传感数据。其中，预设帧号、预设置信度和预设计算资源可基于实际情况进行设置，本申请实施例对此不做具体限定。例如，预设帧号为50，预设置信度为0.8。
通过当前帧传感数据的帧号来确定当前帧传感数据是否为关键帧传感数据，在传感器采集传感数据的过程中，能够间隔一段时间确定当前帧传感数据为关键帧传感数据，从而使用基于全局的第二目标检测模型来对当前帧传感数据进行目标检测，可以在时序上保证目标检测结果的稳定性和准确性，还可以把传感数据时间序列中的第一帧传感数据作为关键帧传感数据，后续的每一帧传感数据均作为非关键帧传感数据。因此，在当前帧传感数据不为关键帧传感数据的一段时间内，使用基于局部区域的第一目标检测模型对候选区域进行目标检测，可以在保证目标检测结果的准确性的同时，减少计算资源的消耗。
在目标对象在上一帧的目标检测信息的目标置信度小于预设置信度，也即上一帧的目标检测信息的置信度较低时，可移动平台基于置信度较低的目标检测信息来规划自身的行驶，容易出现安全事故，无法保证可移动平台安全，因此，在目标对象在上一帧的目标检测信息的目标置信度小于预设置信度时，将当前帧传感数据确定为关键帧传感数据，从而使用基于全局的第二目标检测模型来对当前帧传感数据进行目标检测，可以提高目标检测信息的置信度，使得可移动平台能够基于准确的目标检测信息来来规划自身的行驶，避免出现安全事故，提高可移动平台的安全性。
通过在可移动平台的剩余计算资源大于预设计算资源时，将当前帧传感数据确定为关键帧传感数据，从而使用基于全局的第二目标检测模型来对当前帧传感数据进行目标检测，可以提高目标检测信息的准确性，而在可移动平台的剩余计算资源小于或等于预设计算资源时，将当前帧传感数据确定为非关键帧传感数据，从而使用基于局部区域的第一目标检测模型来对候选区域进行目标检测，可以在保证目标检测结果的准确性的同时，减少计算资源的消耗。
在一实施例中，确定目标对象在上一帧的目标检测信息的目标置信度的方式可以为：获取每个第一候选区域中的目标对象的类别置信度和/或定位置信度；根据每个类别置信度和/或定位置信度，确定目标对象在上一帧的目标检测信息的目标置信度。示例性的，根据每个类别置信度，确定类别置信度的平均值，并将类别置信度的平均值确定为目标对象在上一帧的目标检测信息的目标置信度。或者，根据每个定位置信度，确定定位置信度的平均值，并将定位置信度的平均值确定为目标对象在上一帧的目标检测信息的目标置信度。或者，根据每个类别置信度，确定类别置信度的平均值，并将类别置信度的平均值确定为第一置信度，同时根据每个定位置信度，确定定位置信度的平均值，并将定位置信度的平均值确定为第二置信度，然后确定第一置信度与第二置信度的平均 值，将第一置信度与第二置信度的平均值确定为目标对象在上一帧的目标检测信息的目标置信度。
在一实施例中，获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域；获取目标对象在上一帧的目标检测信息，并基于预设时序预测算法和目标对象在上一帧的目标检测信息，预测目标对象在当前帧的目标检测信息；根据预测的目标对象在当前帧的目标检测信息，确定目标对象在当前帧传感数据中的第四候选区域；根据第二候选区域和第四候选区域，确定目标对象在当前帧的目标检测信息。通过目标对象在上一帧的目标检测信息和候选区域，可以预测得到目标对象在当前帧的候选区域，提高候选区域的质量，可以提高目标检测结果的准确性和算法的鲁棒性。
在一实施例中，也可以根据第二候选区域、第三候选区域和第四候选区域，确定目标对象在当前帧的目标检测信息。还可以根据第二候选区域、第四候选区域和当前帧传感数据，确定目标对象在当前帧的目标检测信息。还可以根据第二候选区域、第三候选区域、第四候选区域和当前帧传感数据，确定目标对象在当前帧的目标检测信息。本申请实施例对此不做具体限定。
上述实施例提供的目标检测方法，通过获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域，然后根据第一候选区域，确定目标对象在当前帧传感数据中的第二候选区域，最后根据第二候选区域，确定目标对象的目标检测信息，由于第二候选区域是基于目标对象在上一帧传感数据中的第一候选区域确定的，因此考虑了当前帧传感数据与上一帧传感数据之间的时序信息，极大的提高了目标检测结果的稳定性和准确性。
请参阅图9，图9是本申请实施例提供的一种目标检测装置的结构示意性框图。
如图9所示，该目标检测装置200包括处理器201和存储器202，处理器201和存储器202通过总线203连接，该总线203比如为I2C(Inter-integrated Circuit)总线。
具体地，处理器201可以是微控制单元(Micro-controller Unit，MCU)、中央处理单元(Central Processing Unit，CPU)或数字信号处理器(Digital Signal Processor，DSP)等。
具体地，存储器202可以是Flash芯片、只读存储器(ROM，Read-Only Memory)磁盘、光盘、U盘或移动硬盘等。
其中，所述处理器201用于运行存储在存储器202中的计算机程序，并在执行所述计算机程序时实现如下步骤：
获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；
根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域；
根据所述第二候选区域，确定所述目标对象的目标检测信息。
所述第一候选区域为多个，所述处理器在实现根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域时，用于实现：
获取每个所述第一候选区域对应的第一目标检测信息；
根据每个所述第一目标检测信息，对多个所述第一候选区域进行过滤，得到至少一个所述第二候选区域。
在一实施例中，所述第一目标检测信息包括所述目标对象的高度和位置坐标，所述第二候选区域中的所述目标对象的所述高度小于或等于预设高度，和/或，所述第二候选区域中的所述目标对象的所述位置坐标位于预设位置坐标范围。
在一实施例中，所述处理器在实现根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域时，用于实现：
获取所述第一候选区域对应的第一目标检测信息；
根据所述第一目标检测信息和预设时序预测算法，预测所述目标对象在当前帧的第二目标检测信息；
根据预测的所述目标对象的第二目标检测信息，确定所述目标对象在所述当前帧传感数据中的第二候选区域。
在一实施例中，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
将所述第二候选区域输入预设的第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
在一实施例中，所述处理器还用于实现以下步骤：
获取多个第一训练样本数据，其中，所述第一训练样本数据包括目标对象在传感数据中的候选区域、标注的目标检测信息和标注的候选区域；
根据所述多个第一训练样本数据对第一神经网络模型进行迭代训练，直到迭代训练后的第一神经网络模型收敛，得到所述第一目标检测模型。
在一实施例中，所述处理器还用于实现以下步骤：
显示所述目标候选区域。
在一实施例中，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域；
根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息。
在一实施例中，所述处理器在实现根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域时，用于实现：
根据传感器的预设安装信息，确定所述目标对象在所述传感器的坐标系下的预设位置坐标；
根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域。
在一实施例中，所述处理器在实现根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域时，用于实现：
确定所述预设位置坐标与多个预设位置坐标增益中的每个预设位置坐标增益之间的和值和/或差值，得到多个候选位置坐标；
根据所述多个候选位置坐标和所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的多个第三候选区域。
在一实施例中，所述处理器在实现根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息时，用于实现：
将所述第二候选区域和所述第三候选区域输入第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
在一实施例中，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
在一实施例中，所述处理器在实现根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
将所述第二候选区域输入预设的第一目标检测模型，得到第三目标检测信息和所述目标对象在所述当前帧传感数据中的第四候选区域；
将所述当前帧传感数据输入预设的第二目标检测模型，得到第四目标检测信息和所述目标对象在所述当前帧传感数据中的第五候选区域；
根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息；
根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域。
在一实施例中，所述处理器在实现根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息时，用于实现：
确定所述第三目标检测信息与所述第四目标检测信息之间的匹配度；
若所述匹配度大于或等于预设匹配度，则对所述第三目标检测信息和所述第四目标检测信息进行融合，得到所述目标对象的目标检测信息。
在一实施例中，所述处理器还用于实现以下步骤：
若所述匹配度小于预设匹配度，则将所述第四目标检测信息确定为所述目标对象的目标检测信息。
在一实施例中，所述第四候选区域和所述第五候选区域均为多个，所述处理器在实现根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域时，用于实现：
确定每个所述第四候选区域与每个所述第五候选区域之间的相似度；
根据所述相似度从多个所述第四候选区域和多个所述第五候选区域中确定目标候选区域对，所述目标候选区域对包括一个所述第四候选区域和一个所述第五候选区域；
将所述目标候选区域对中的所述第四候选区域和/或第五候选区域确定为所述目标候选区域。
在一实施例中，所述目标候选区域对中的所述第四候选区域与所述第五候选区域的相似度大于预设相似度。
在一实施例中，所述处理器还用于实现以下步骤：
获取多个第二训练样本数据，其中，所述第二训练样本数据包括传感数据，标注的目标检测信息和标注的候选区域；
根据所述多个第二训练样本数据对第二神经网络模型进行迭代训练，直到迭代训练后的第二神经网络模型收敛，得到所述第二目标检测模型。
在一实施例中，运行所述第一目标检测模型所需的第一计算资源小于运行所述第二目标检测模型所需的第二计算资源。
在一实施例中，所述处理器还用于实现以下步骤：
若所述当前帧传感数据为关键帧传感数据，则根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
在一实施例中，所述处理器还用于实现以下步骤：
获取所述当前帧传感数据的帧号、所述目标对象在上一帧的目标检测信息的目标置信度和/或剩余计算资源；
根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据。
在一实施例中，所述处理器在实现根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据时，用于实现：
若所述当前帧传感数据的帧号为预设帧号的整数倍、所述目标置信度小于预设置信度和/或所述剩余计算资源大于预设计算资源，则确定所述当前帧传感数据为关键帧传感数据。
在一实施例中，所述处理器还用于实现以下步骤：
获取每个所述第一候选区域中的目标对象的类别置信度和/或定位置信度；
根据每个所述类别置信度和/或定位置信度，确定所述目标对象在上一帧的目标检测信息的目标置信度。
在一实施例中，传感器包括视觉传感器和雷达装置。
需要说明的是，所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的目标检测装置的具体工作过程，可以参考前述目标检测方法实施例中的对应过程，在此不再赘述。
请参阅图10，图10是本申请实施例提供的一种可移动平台的结构示意性框图。
如图10所示，可移动平台300包括平台本体310、动力系统320、传感器330和目标检测装置340，动力系统320、传感器330和目标检测装置340设于平台本体310上，动力系统320用于为可移动平台300提供移动动力，传感器330用于采集传感数据，目标检测装置340用于确定目标对象的目标检测信息以及还用于控制可移动平台300。可移动平台300包括无人机、机器人、无人船和无人驾驶汽车等。
需要说明的是，所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的可移动平台的具体工作过程，可以参考前述目标检测方法 实施例中的对应过程，在此不再赘述。
本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序中包括程序指令，所述处理器执行所述程序指令，实现上述实施例提供的目标检测方法的步骤。
其中，所述计算机可读存储介质可以是前述任一实施例所述的可移动平台的内部存储单元，例如所述可移动平台的硬盘或内存。所述计算机可读存储介质也可以是所述可移动平台的外部存储设备，例如所述可移动平台上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。
应当理解，在此本申请说明书中所使用的术语仅仅是出于描述特定实施例的目的而并不意在限制本申请。如在本申请说明书和所附权利要求书中所使用的那样，除非上下文清楚地指明其它情况，否则单数形式的“一”、“一个”及“该”意在包括复数形式。
还应当理解，在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以权利要求的保护范围为准。

Claims (50)

  1. 一种目标检测方法，其特征在于，包括：
    获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；
    根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域；
    根据所述第二候选区域，确定所述目标对象的目标检测信息。
  2. 根据权利要求1所述的目标检测方法，其特征在于，所述第一候选区域为多个，所述根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域，包括：
    获取每个所述第一候选区域对应的第一目标检测信息；
    根据每个所述第一目标检测信息，对多个所述第一候选区域进行过滤，得到至少一个所述第二候选区域。
  3. 根据权利要求2所述的目标检测方法，其特征在于，所述第一目标检测信息包括所述目标对象的高度和位置坐标，所述第二候选区域中的所述目标对象的所述高度小于或等于预设高度，和/或，所述第二候选区域中的所述目标对象的所述位置坐标位于预设位置坐标范围。
  4. 根据权利要求1所述的目标检测方法，其特征在于，所述根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域，包括：
    获取所述第一候选区域对应的第一目标检测信息；
    根据所述第一目标检测信息和预设时序预测算法，预测所述目标对象在当前帧的第二目标检测信息；
    根据预测的所述目标对象的第二目标检测信息，确定所述目标对象在所述当前帧传感数据中的第二候选区域。
  5. 根据权利要求1所述的目标检测方法，其特征在于，所述根据所述第二候选区域，确定所述目标对象的目标检测信息，包括：
    将所述第二候选区域输入预设的第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
  6. 根据权利要求5所述的目标检测方法，其特征在于，所述方法还包括：
    获取多个第一训练样本数据，其中，所述第一训练样本数据包括目标对象在传感数据中的候选区域、标注的目标检测信息和标注的候选区域；
    根据所述多个第一训练样本数据对第一神经网络模型进行迭代训练，直到 迭代训练后的第一神经网络模型收敛，得到所述第一目标检测模型。
  7. 根据权利要求5所述的目标检测方法，其特征在于，所述方法还包括：
    显示所述目标候选区域。
  8. 根据权利要求1所述的目标检测方法，其特征在于，所述根据所述第二候选区域，确定所述目标对象的目标检测信息，包括：
    根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域；
    根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息。
  9. 根据权利要求8所述的目标检测方法，其特征在于，所述根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域，包括：
    根据传感器的预设安装信息，确定所述目标对象在所述传感器的坐标系下的预设位置坐标；
    根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域。
  10. 根据权利要求9所述的目标检测方法，其特征在于，所述根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域，包括：
    确定所述预设位置坐标与多个预设位置坐标增益中的每个预设位置坐标增益之间的和值和/或差值，得到多个候选位置坐标；
    根据所述多个候选位置坐标和所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的多个第三候选区域。
  11. 根据权利要求8所述的目标检测方法，其特征在于，所述根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息，包括：
    将所述第二候选区域和所述第三候选区域输入第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
  12. 根据权利要求1-11中任一项所述的目标检测方法，其特征在于，所述根据所述第二候选区域，确定所述目标对象的目标检测信息，包括：
    根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
  13. 根据权利要求12所述的目标检测方法，其特征在于，所述根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息，包括：
    将所述第二候选区域输入预设的第一目标检测模型，得到第三目标检测信息和所述目标对象在所述当前帧传感数据中的第四候选区域；
    将所述当前帧传感数据输入预设的第二目标检测模型，得到第四目标检测信息和所述目标对象在所述当前帧传感数据中的第五候选区域；
    根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息；
    根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域。
  14. 根据权利要求13所述的目标检测方法，其特征在于，所述根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息，包括：
    确定所述第三目标检测信息与所述第四目标检测信息之间的匹配度；
    若所述匹配度大于或等于预设匹配度，则对所述第三目标检测信息和所述第四目标检测信息进行融合，得到所述目标对象的目标检测信息。
  15. 根据权利要求14所述的目标检测方法，其特征在于，所述方法还包括：
    若所述匹配度小于预设匹配度，则将所述第四目标检测信息确定为所述目标对象的目标检测信息。
  16. 根据权利要求13所述的目标检测方法，其特征在于，所述第四候选区域和所述第五候选区域均为多个，所述根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域，包括：
    确定每个所述第四候选区域与每个所述第五候选区域之间的相似度；
    根据所述相似度从多个所述第四候选区域和多个所述第五候选区域中确定目标候选区域对，所述目标候选区域对包括一个所述第四候选区域和一个所述第五候选区域；
    将所述目标候选区域对中的所述第四候选区域和/或第五候选区域确定为所述目标候选区域。
  17. 根据权利要求16所述的目标检测方法，其特征在于，所述目标候选区域对中的所述第四候选区域与所述第五候选区域的相似度大于预设相似度。
  18. 根据权利要求13所述的目标检测方法，其特征在于，所述方法还包括：
    获取多个第二训练样本数据，其中，所述第二训练样本数据包括传感数据， 标注的目标检测信息和标注的候选区域；
    根据所述多个第二训练样本数据对第二神经网络模型进行迭代训练，直到迭代训练后的第二神经网络模型收敛，得到所述第二目标检测模型。
  19. 根据权利要求13所述的目标检测方法，其特征在于，运行所述第一目标检测模型所需的第一计算资源小于运行所述第二目标检测模型所需的第二计算资源。
  20. 根据权利要求1-11中任一项所述的目标检测方法，其特征在于，所述方法还包括：
    若所述当前帧传感数据为关键帧传感数据，则根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
  21. 根据权利要求1-11中任一项所述的目标检测方法，其特征在于，所述方法还包括：
    获取所述当前帧传感数据的帧号、所述目标对象在上一帧的目标检测信息的目标置信度和/或剩余计算资源；
    根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据。
  22. 根据权利要求21所述的目标检测方法，其特征在于，所述根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据，包括：
    若所述当前帧传感数据的帧号为预设帧号的整数倍、所述目标置信度小于预设置信度和/或所述剩余计算资源大于预设计算资源，则确定所述当前帧传感数据为关键帧传感数据。
  23. 根据权利要求21所述的目标检测方法，其特征在于，所述方法还包括：
    获取每个所述第一候选区域中的目标对象的类别置信度和/或定位置信度；
    根据每个所述类别置信度和/或定位置信度，确定所述目标对象在上一帧的目标检测信息的目标置信度。
  24. 根据权利要求1-11中任一项所述的目标检测方法，其特征在于，传感器包括视觉传感器和雷达装置。
  25. 一种目标检测装置，其特征在于，所述目标检测装置包括存储器和处理器；
    所述存储器用于存储计算机程序；
    所述处理器，用于执行所述计算机程序并在执行所述计算机程序时，实现 如下步骤：
    获取当前帧传感数据和目标对象在上一帧传感数据中的第一候选区域；
    根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域；
    根据所述第二候选区域，确定所述目标对象的目标检测信息。
  26. 根据权利要求25所述的目标检测装置，其特征在于，所述第一候选区域为多个，所述处理器在实现根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域时，用于实现：
    获取每个所述第一候选区域对应的第一目标检测信息；
    根据每个所述第一目标检测信息，对多个所述第一候选区域进行过滤，得到至少一个所述第二候选区域。
  27. 根据权利要求26所述的目标检测装置，其特征在于，所述第一目标检测信息包括所述目标对象的高度和位置坐标，所述第二候选区域中的所述目标对象的所述高度小于或等于预设高度，和/或，所述第二候选区域中的所述目标对象的所述位置坐标位于预设位置坐标范围。
  28. 根据权利要求25所述的目标检测装置，其特征在于，所述处理器在实现根据所述第一候选区域，确定所述目标对象在所述当前帧传感数据中的第二候选区域时，用于实现：
    获取所述第一候选区域对应的第一目标检测信息；
    根据所述第一目标检测信息和预设时序预测算法，预测所述目标对象在当前帧的第二目标检测信息；
    根据预测的所述目标对象的第二目标检测信息，确定所述目标对象在所述当前帧传感数据中的第二候选区域。
  29. 根据权利要求25所述的目标检测装置，其特征在于，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
    将所述第二候选区域输入预设的第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
  30. 根据权利要求29所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    获取多个第一训练样本数据，其中，所述第一训练样本数据包括目标对象在传感数据中的候选区域、标注的目标检测信息和标注的候选区域；
    根据所述多个第一训练样本数据对第一神经网络模型进行迭代训练，直到 迭代训练后的第一神经网络模型收敛，得到所述第一目标检测模型。
  31. 根据权利要求29所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    显示所述目标候选区域。
  32. 根据权利要求25所述的目标检测装置，其特征在于，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
    根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域；
    根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息。
  33. 根据权利要求32所述的目标检测装置，其特征在于，所述处理器在实现根据传感器的预设安装信息，确定所述目标对象在所述当前帧传感数据中的第三候选区域时，用于实现：
    根据传感器的预设安装信息，确定所述目标对象在所述传感器的坐标系下的预设位置坐标；
    根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域。
  34. 根据权利要求33所述的目标检测装置，其特征在于，所述处理器在实现根据所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的第三候选区域时，用于实现：
    确定所述预设位置坐标与多个预设位置坐标增益中的每个预设位置坐标增益之间的和值和/或差值，得到多个候选位置坐标；
    根据所述多个候选位置坐标和所述预设位置坐标，确定所述目标对象在所述当前帧传感数据中的多个第三候选区域。
  35. 根据权利要求32所述的目标检测装置，其特征在于，所述处理器在实现根据所述第二候选区域和所述第三候选区域，确定所述目标对象的目标检测信息时，用于实现：
    将所述第二候选区域和所述第三候选区域输入第一目标检测模型，得到所述目标对象的目标检测信息和所述目标对象在所述当前帧传感数据中的目标候选区域。
  36. 根据权利要求25-35中任一项所述的目标检测装置，其特征在于，所述处理器在实现根据所述第二候选区域，确定所述目标对象的目标检测信息时， 用于实现：
    根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
  37. 根据权利要求36所述的目标检测装置，其特征在于，所述处理器在实现根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息时，用于实现：
    将所述第二候选区域输入预设的第一目标检测模型，得到第三目标检测信息和所述目标对象在所述当前帧传感数据中的第四候选区域；
    将所述当前帧传感数据输入预设的第二目标检测模型，得到第四目标检测信息和所述目标对象在所述当前帧传感数据中的第五候选区域；
    根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息；
    根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域。
  38. 根据权利要求37所述的目标检测装置，其特征在于，所述处理器在实现根据所述第三目标检测信息和所述第四目标检测信息，确定所述目标对象的目标检测信息时，用于实现：
    确定所述第三目标检测信息与所述第四目标检测信息之间的匹配度；
    若所述匹配度大于或等于预设匹配度，则对所述第三目标检测信息和所述第四目标检测信息进行融合，得到所述目标对象的目标检测信息。
  39. 根据权利要求38所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    若所述匹配度小于预设匹配度，则将所述第四目标检测信息确定为所述目标对象的目标检测信息。
  40. 根据权利要求37所述的目标检测装置，其特征在于，所述第四候选区域和所述第五候选区域均为多个，所述处理器在实现根据所述第四候选区域和所述第五候选区域，确定所述目标对象在所述当前帧传感数据中的目标候选区域时，用于实现：
    确定每个所述第四候选区域与每个所述第五候选区域之间的相似度；
    根据所述相似度从多个所述第四候选区域和多个所述第五候选区域中确定目标候选区域对，所述目标候选区域对包括一个所述第四候选区域和一个所述第五候选区域；
    将所述目标候选区域对中的所述第四候选区域和/或第五候选区域确定为所述目标候选区域。
  41. 根据权利要求40所述的目标检测装置，其特征在于，所述目标候选区域对中的所述第四候选区域与所述第五候选区域的相似度大于预设相似度。
  42. 根据权利要求37所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    获取多个第二训练样本数据，其中，所述第二训练样本数据包括传感数据，标注的目标检测信息和标注的候选区域；
    根据所述多个第二训练样本数据对第二神经网络模型进行迭代训练，直到迭代训练后的第二神经网络模型收敛，得到所述第二目标检测模型。
  43. 根据权利要求37所述的目标检测装置，其特征在于，运行所述第一目标检测模型所需的第一计算资源小于运行所述第二目标检测模型所需的第二计算资源。
  44. 根据权利要求25-35中任一项所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    若所述当前帧传感数据为关键帧传感数据，则根据所述当前帧传感数据和所述第二候选区域，确定所述目标对象的目标检测信息。
  45. 根据权利要求25-35中任一项所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    获取所述当前帧传感数据的帧号、所述目标对象在上一帧的目标检测信息的目标置信度和/或剩余计算资源；
    根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据。
  46. 根据权利要求45所述的目标检测装置，其特征在于，所述处理器在实现根据所述当前帧传感数据的帧号、所述目标置信度和/或剩余计算资源，确定所述当前帧传感数据是否为关键帧传感数据时，用于实现：
    若所述当前帧传感数据的帧号为预设帧号的整数倍、所述目标置信度小于预设置信度和/或所述剩余计算资源大于预设计算资源，则确定所述当前帧传感数据为关键帧传感数据。
  47. 根据权利要求45所述的目标检测装置，其特征在于，所述处理器还用于实现以下步骤：
    获取每个所述第一候选区域中的目标对象的类别置信度和/或定位置信度；
    根据每个所述类别置信度和/或定位置信度，确定所述目标对象在上一帧的目标检测信息的目标置信度。
  48. 根据权利要求25-35中任一项所述的目标检测装置，其特征在于，传感器包括视觉传感器和雷达装置。
  49. 一种可移动平台，其特征在于，包括：
    平台本体；
    动力系统，设于所述平台本体上，用于为所述可移动平台提供移动动力；
    传感器，设于所述平台本体上，用于采集传感数据；
    权利要求25-48中任一项所述目标检测装置，设于所述平台本体内，用于确定目标对象的目标检测信息以及还用于控制所述可移动平台。
  50. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时使所述处理器在实现如权利要求1-24中任一项所述的目标检测方法的步骤。
PCT/CN2020/139043 2020-12-24 2020-12-24 目标检测方法、装置、可移动平台及计算机可读存储介质 WO2022133911A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/139043 WO2022133911A1 (zh) 2020-12-24 2020-12-24 目标检测方法、装置、可移动平台及计算机可读存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/139043 WO2022133911A1 (zh) 2020-12-24 2020-12-24 目标检测方法、装置、可移动平台及计算机可读存储介质

Publications (1)

Publication Number Publication Date
WO2022133911A1 true WO2022133911A1 (zh) 2022-06-30

Family

ID=82157198

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/139043 WO2022133911A1 (zh) 2020-12-24 2020-12-24 目标检测方法、装置、可移动平台及计算机可读存储介质

Country Status (1)

Country Link
WO (1) WO2022133911A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180137647A1 (en) * 2016-11-15 2018-05-17 Samsung Electronics Co., Ltd. Object detection method and apparatus based on dynamic vision sensor
CN110555405A (zh) * 2019-08-30 2019-12-10 北京迈格威科技有限公司 目标跟踪方法及装置、存储介质和电子设备
CN111427979A (zh) * 2020-01-15 2020-07-17 深圳市镭神智能系统有限公司 基于激光雷达的动态地图构建方法、系统及介质
CN111428567A (zh) * 2020-02-26 2020-07-17 沈阳大学 一种基于仿射多任务回归的行人跟踪系统及方法
CN111428566A (zh) * 2020-02-26 2020-07-17 沈阳大学 一种形变目标跟踪系统及方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180137647A1 (en) * 2016-11-15 2018-05-17 Samsung Electronics Co., Ltd. Object detection method and apparatus based on dynamic vision sensor
CN110555405A (zh) * 2019-08-30 2019-12-10 北京迈格威科技有限公司 目标跟踪方法及装置、存储介质和电子设备
CN111427979A (zh) * 2020-01-15 2020-07-17 深圳市镭神智能系统有限公司 基于激光雷达的动态地图构建方法、系统及介质
CN111428567A (zh) * 2020-02-26 2020-07-17 沈阳大学 一种基于仿射多任务回归的行人跟踪系统及方法
CN111428566A (zh) * 2020-02-26 2020-07-17 沈阳大学 一种形变目标跟踪系统及方法

Similar Documents

Publication Publication Date Title
US11749124B2 (en) User interaction with an autonomous unmanned aerial vehicle
JP7086111B2 (ja) 自動運転車のｌｉｄａｒ測位に用いられるディープラーニングに基づく特徴抽出方法
US20200026282A1 (en) Lane/object detection and tracking perception system for autonomous vehicles
JP7060625B2 (ja) 自動運転車において３ｄｃｎｎネットワークを用いてソリューション推断を行うｌｉｄａｒ測位
CN111874006B (zh) 路线规划处理方法和装置
US11164369B2 (en) Methods and systems for constructing map data using poisson surface reconstruction
CN112740268B (zh) 目标检测方法和装置
US11657532B2 (en) Surface profile estimation and bump detection for autonomous machine applications
WO2021217420A1 (zh) 车道线跟踪方法和装置
EP3674971B1 (en) Method and system for training machine learning algorithm to detect objects at distance
JP2021515178A (ja) 自動運転車両においてｒｎｎとｌｓｔｍを用いて時間平滑化を行うｌｉｄａｒ測位
WO2020186444A1 (zh) 物体检测方法、电子设备与计算机存储介质
WO2020181418A1 (en) Techniques for collaborative map construction between unmanned aerial vehicle and ground vehicle
EP3729333A1 (en) Local sensing based autonomous navigation, and associated systems and methods
EP4222035A1 (en) Methods and systems for performing outlet inference by an autonomous vehicle to determine feasible paths through an intersection
CN114973050A (zh) 自动驾驶应用中深度神经网络感知的地面实况数据生成
Lei et al. Automated Lane Change Behavior Prediction and Environmental Perception Based on SLAM Technology
WO2024036984A1 (zh) 目标定位方法及相关系统、存储介质
WO2022133911A1 (zh) 目标检测方法、装置、可移动平台及计算机可读存储介质
US11645762B2 (en) Obstacle detection
US20240151855A1 (en) Lidar-based object tracking
US11884268B2 (en) Motion planning in curvilinear coordinates for autonomous vehicles
Zhang et al. Vision-based uav positioning method assisted by relative attitude classification
WO2022160101A1 (zh) 朝向估计方法、装置、可移动平台及可读存储介质
US20230237793A1 (en) False track mitigation in object detection systems

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20966483

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20966483

Country of ref document: EP

Kind code of ref document: A1