CN116858253A - A lightweight predictive navigation method and system suitable for indoor environments - Google Patents

Abstract

The invention discloses a lightweight predictive navigation method and system suitable for indoor environments. The environment prediction combines the obtained sensory information and motion information to generate future environment prediction and mapping, which is expressed as an occupied grid sequence; Motion planning uses a grid sequence with time attributes as a search space, formulates the trajectory solution as a nonlinear optimization problem, and solves the nonlinear optimization problem to obtain the trajectory; in a unified environment representation form, obtains sensing data and vehicle's own motion data (Speed, coordinate) partially decouples the dynamic and static environment. The static environment obtains the future position through coordinate transformation and mapping. The dynamic environment predicts the raster map of several steps in the future by extracting scene motion feature information, and then directly uses the future raster. At the same time Process dynamic and static obstacles to improve the efficiency and effectiveness of trajectory optimization; the method proposed by the present invention can improve real-time performance and safety while reducing computing requirements.

Description

A lightweight predictive navigation method and system suitable for indoor environments

Technical field

The invention belongs to the field of intelligent navigation technology, and specifically relates to a lightweight predictive navigation method and system suitable for indoor environments.

Background technique

To complete specific tasks, robots require precise and safe navigation to reach their destination without collisions. This navigation is achieved by implementing navigation algorithms. In a dynamic environment, the algorithm consists of two main components: prediction and planning. The prediction component provides information about local obstacle distribution, while the planning component generates safe and seamless trajectories based on the prediction information to adapt to dynamic changes. Although the researchers studied separate implementations of both functions, their performance was found to be insufficient in lightweight, unstructured indoor dynamic environments. Due to space congestion, high obstacle density, small size, and lack of unified representation, obstacle information is lost. In addition, the limited space constrains the agent, and its poor real-time performance prevents it from performing evasive maneuvers in time, leading to collisions.

Most existing methods have high requirements on sensor systems and are time-consuming, making it difficult to meet the real-time requirements of agent planning. Additionally, these methods employ object tracking to obtain future trajectories of dynamic obstacles, which are then used as inputs to the planning module. However, many methods rely on large-scale models, resulting in poor real-time performance. This can be attributed to limitations in cost and size of robots, making it challenging to implement predictive methods that require excessive resources. Furthermore, the lack of a unified representation between prediction and planning results in the loss of essential information during the conversion process of multiple input and output formats. Additionally, planning algorithms often lack foresight into the future behavior of obstacles, which can cause the planning algorithm to get bogged down, resulting in trajectory deviations or discontinuities. Most existing prediction methods rely heavily on multi-line lidar, which is expensive and generates large-scale point clouds, adding additional computational burden to indoor agent navigation. Traditional planning algorithms require high-quality prediction information but have low robustness, resulting in limited practical applicability.

Contents of the invention

In order to solve the problems existing in the prior art, the present invention provides a lightweight predictive navigation method suitable for indoor environments. It adopts a hierarchical prediction-planning method to obtain sensing data and vehicle own motion in a unified environment representation form. The data (velocity, coordinates) part decouples the dynamic and static environment. The static environment obtains the future position through coordinate transformation and mapping. The dynamic environment predicts the raster map of several steps in the future by extracting scene motion feature information, and then directly uses the future raster. It also handles dynamic and static obstacles to improve the efficiency and effectiveness of trajectory optimization.

In order to achieve the above objectives, the technical solution adopted by the present invention is: a lightweight predictive navigation method suitable for indoor environments, which includes the following steps:

By performing coordinate conversion of the sensor measurement values on the pose information, speed information and environmental representation of the intelligent body obtained by the multi-sensor system, the environmental variables are projected into the coordinate system of the future intelligent body;

Consider the environmental variables in the coordinate system where the future agent is located, the propulsion of dynamic obstacles and the occupation of static obstacles in the environment, capture the interaction between the two, and decouple the dynamic components and static components to extract the future Probabilistic occupancy raster plot of latent variable characteristics;

A variational autoencoder is used to build an OGM occupancy raster map prediction generation module. The latent variable characteristics of the future probability occupancy raster map are used as input, and the decoder outputs the predicted occupancy raster map sequence;

Each raster map is used sequentially as a feasible space for sampling candidate trajectories;

Generate candidate trajectories in the feasible space and obtain initial trajectories that satisfy kinematic constraints and obstacle constraints;

Graph optimization methods are used to generate smoother and safer trajectories based on the initial trajectory.

Furthermore, by performing coordinate conversion of the sensor measurement values on the pose information, speed information and environmental representation of the intelligent body obtained by the multi-sensor system, projecting the environmental variables into the coordinate system of the future intelligent body includes:

First, calculate the current obstacle occupation position based on the measurement value of the lidar, and obtain the obstacle occupation position in the coordinate system where the current agent is located;

According to the conversion relationship between the current and next poses of the agent and the current pose information of the agent, the pose of the agent at the next moment is inferred; the current obstacle distribution is projected to the pose of the agent after a set period of time. coordinate system to obtain the projection of environmental variables in the coordinate system where the future agent is located.

Furthermore, consider the propulsion of dynamic obstacles and the occupation of static obstacles in the environment, capture the interaction between the two, decouple the dynamic components and static components, and extract the specific latent variable characteristics of the future probability occupation raster map. Includes the following steps:

The entire grid map information is fed into the convolutional long and short-term memory network module. The convolutional long and short-term memory network module takes the one-dimensional occupied grid map as input and uses a convolution kernel of size 3 to generate dynamic grid features;

The Bayesian generation method is used to predict the local map of future static obstacles, and at the same time filter out the dynamic information to obtain the local map information of the static obstacles;

The extracted dynamic grid features and local map information of static obstacles are used as inputs to the OGM occupancy grid map prediction generation module.

Further, the OGM occupied raster map prediction generation module includes an encoder and a decoder. The encoder is composed of two convolutional layers and two residual layers. The parameters of the two residual layers are consistent; the decoder is composed of two residual layers. It consists of a difference layer and two deconvolution layers, and the residual layer structure is consistent with the structure in the encoder.

Further, when each raster map is used sequentially as a feasible space for sampling candidate trajectories, specifically, based on the extended sequence of occupied raster maps, the occupied space in the prediction sequence is derived through the position transformation relationship, and in the set time period Within, the agent moves uniformly, and the center of mass is the center of the circle. The feasible space representation of each frame is obtained through the area that the agent can reach. The state within the feasible interval obeys strict speed interval constraints and collision constraints.

Further, when generating candidate trajectories in the feasible space and obtaining the initial trajectory that satisfies kinematic constraints and obstacle constraints, a random sampling method is used to select frames in the feasible space, and the raster image sequence is sampled one by one, and the spatial representation of the sampling is For the unoccupied area in each graph, the search space is obtained on the basis of satisfying the kinematic constraints of the agent and ensuring reachability during sampling; during sampling, if the sampling points between adjacent frames can be moved without any collision , and determine whether the connected edge satisfies the dynamics and kinematic constraints of the agent. If so, connect the sampling points together to form an edge; traverse the node list to identify the node with the lowest cost, and start backtracking The process obtains an initial trajectory that satisfies kinematic constraints and obstacle constraints.

Further, using graph optimization methods to generate smoother and safer trajectories based on the initial trajectory includes:

The generation of smoother and safer trajectories using graph optimization methods is formulated as a nonlinear optimization problem:

Among them, L({P _t }|{m _t:t+β }) represents the optimization goal, {m _t:t+β } represents the grid map sequence, C _ego represents the outer contour of the self-car, and C _obs represents the outer contour of the obstacle. contour, represents the maximum speed and minimum speed that can be achieved, ω _min , ω _{max} represents the maximum angular velocity and minimum angular velocity that can be achieved, |a _v |≤a _{max} , which means that the linear acceleration is less than the maximum linear acceleration, Indicates that the angular acceleration is less than the maximum angular acceleration;

The hard constraints are converted into penalties within the objective function, and the hard constraints are regarded as soft constraints; the loop is iteratively optimized until the optimized trajectory satisfies the kinematic constraints and obstacle constraints or reaches the upper limit of the iteration, and the final planned path is obtained by solution.

Based on the technical concept of the method, the present invention also provides a lightweight predictive navigation system suitable for indoor environments, including a pose prediction module, a dynamic and static feature extraction module, an occupancy grid sequence prediction module, and a feasible space acquisition module. , initial trajectory acquisition module and trajectory optimization module;

The pose prediction module performs coordinate conversion of sensory measurement values on the pose information, speed information and environmental representation of the intelligent body obtained by the multi-sensor system, and projects the environmental variables into the coordinate system of the future intelligent body;

The dynamic and static feature extraction module is used to consider the environmental variables in the coordinate system where the future agent is located, the propulsion of dynamic obstacles and the occupation of static obstacles in the environment, capture the interaction between the two, and combine the dynamic components with Decouple static components to extract latent variable features of future probability occupation raster maps;

The occupancy raster map sequence prediction module uses a variational autoencoder to build an OGM occupancy raster map prediction generation module, taking the latent variable characteristics of the future probability occupancy raster map as input, and the decoder outputs the predicted occupancy raster map sequence;

The feasible space acquisition module is used to sequentially use each raster map as a feasible space for sampling candidate trajectories;

The initial trajectory acquisition module is used to generate candidate trajectories in the feasible space and obtain initial trajectories that satisfy kinematic constraints and obstacle constraints;

The trajectory optimization module uses graph optimization methods to generate smoother and safer trajectories based on the initial trajectory.

In addition, a computer device is provided, including a processor and a memory. The memory is used to store a computer executable program. The processor reads the computer executable program from the memory and executes it. When the processor executes the program, the invention can be realized. The described lightweight predictive navigation method suitable for indoor environments.

The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the lightweight predictive navigation suitable for indoor environments according to the present invention can be realized. method.

Compared with the prior art, the present invention at least has the following beneficial effects:

The present invention performs environmental prediction based on deep learning, is conducive to self-made data sets, and is suitable for different types of indoor environments. As the collected data of the environment increases, better prediction results can be obtained; a lightweight, fast and universal framework is proposed, which can have good applications in indoor mobile robots and driverless vehicles, with low resource consumption and real-time performance. It has high performance and solves the prediction problem of single-line laser; the movement of the agent is taken into account in the prediction work. This kind of movement information is necessary and helps to distinguish dynamic and static obstacles and provides prior information; dynamic obstacles are determined by the time series grid. The occupancy status indicates that the planning can handle dynamic and static obstacles at the same time to ensure safe passage of the trajectory in the presence of multiple obstacles.

Description of the drawings

Figure 1 is a technical framework diagram of the present invention.

Figure 2 is a framework diagram of the prediction part.

Figure 3 is a structural diagram of the variational autoencoder network used for prediction.

Figure 4 is a partial framework diagram of the planning.

Detailed ways

Exemplary examples of the present application are explained in detail below with reference to the accompanying drawings and specific implementations, which include various details of the embodiments of the present application to facilitate understanding. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, those skilled in the art will make modifications to various equivalent forms of the present invention and all fall within the scope of the appended claims of this application. limited scope.

Real-time navigation is an important component of an agent's ability to drive safely in complex dynamic environments, including prediction and motion planning. However, most existing methods require high-quality sensor systems and huge computational consumption, which is contrary to real-time requirements. Furthermore, these methods lack a unified representation framework for prediction and planning, resulting in accuracy loss during representation conversion. The present invention proposes a predictive planning framework based on occupancy grid diagrams. The prediction module combines sensory and motion information to handle dynamic and static obstacles and generates a sequence of occupancy grid maps. The planning module samples, selects and optimizes candidate trajectories. Experimental results in multiple scenarios show that the proposed framework can improve real-time performance and security while reducing computational requirements compared to conventional methods.

Figure 1 is a schematic flowchart of the predictive navigation method based on deep learning of the present invention. The present invention is suitable for complex and dynamic environments and has universality for indoor and outdoor environments. The purpose of the invention is to generate a series of future states containing pose and velocity information for an intelligent agent from measurements obtained from sensors. This invention takes the sensor's representation of the current environment and the expected posture of the next time step as input, including initial environment information and potential motion variables, and the expected output is a feasible motion trajectory. The invention is divided into two interrelated parts: environment prediction and motion planning. Environment prediction combines the obtained sensory information with motion information to generate future environment predictions and mappings, expressed as occupied grid sequence; motion planning uses the grid sequence with time attributes as the search space and formulates the trajectory solution It is a nonlinear optimization problem, and the trajectory is obtained by solving the nonlinear optimization problem. The details of the invention are as follows:

Step 1: During the actual operation of the agent, the local coordinate system where the collected lidar data is located will continue to change with the movement of the agent, which will affect the prediction of the occupied raster map. Motion compensation is divided into two steps, namely body pose prediction and coordinate conversion. In the process of pose prediction, the environmental variables are projected into the coordinate system of the future agent by performing coordinate transformation on the pose information, speed information and environmental representation of the intelligent body obtained by the multi-sensor system. The specific details are as follows:

The present invention can use any sensor that can construct a probability grid map. As an example, a low-cost sensor such as a single-line laser is used for illustration. Assume that the measurement value of lidar is M = [d _t , k _t * δ], where d _t is the distance measured by lidar at time t, k _t is the ranging sequence of lidar, and δ is the resolution of lidar, Expressed in terms of angles.

(1) First, calculate the current occupied position of the obstacle based on the measurement value of the lidar. In the coordinate system where the current agent is located, the occupied position of the obstacle is:

[x,y]＝d _t [cos(k _t *δ),sin(k _t *δ)] ^T

According to the inertial measurement unit, the linear velocity V _t , angular velocity ω _t , linear acceleration a _vt and angular acceleration a _ωt of the agent at time t can be obtained.

(2) According to the current pose information of the agent [x _ego , y _ego , θ ₁ ] given by the positioning module, the present invention assumes that the agent performs uniform linear motion within the time interval Δt of 0.1s, and can infer the next The pose of the agent at the moment [x′ _ego ,y′ _ego ,θ ₂ ]. Among them, x, y, and θ respectively represent the horizontal and vertical coordinates and orientation of the location of the agent.

The transformation relationship between the current and next moment poses of the agent can be expressed as follows:

By projecting the current obstacle distribution into the coordinate system of the agent's pose after Δt, the corresponding coordinates [x ^′ , y ^′ ] of the obstacle occupation point can be obtained. The calculation method is as follows:

Among them, [x ^′ , y ^′ ] is the coordinate system of the pose of the agent after Δt, and the corresponding coordinates of the obstacle occupation point can be obtained. [x _ego , y _ego ] is the coordinate system of the intelligent agent in the current global coordinate system. The position of the body, θ ₁ is the orientation, [x ^′ _ego , y ^′ _ego ] is the position of the agent in the global coordinate system after Δt time, θ ₂ is the orientation; and then we get the position of the environmental variables in the future where the agent is Projection in coordinate system.

Step 2: Extract dynamic and static feature information. The process of extracting dynamic and static feature information simultaneously considers the propulsion of dynamic obstacles and the occupancy of static obstacles in the environment, implicitly captures the interaction between the two, and partially decouples the dynamic and static components to extract future probability occupancy Latent variable characteristics of raster plots.

(1) In order to predict dynamic grid information, the entire grid map information is fed into the convolutional long short-term memory network module. The convolutional long short-term memory network module takes a one-dimensional occupancy grid map as input and uses a convolution kernel of size 3 to generate 32-dimensional dynamic grid features.

(2) Regarding the static obstacle information, the Bayesian generation method is used to predict the local map of the future static obstacles, and at the same time filter out the dynamic information to obtain the local map information of the static obstacles. This approach allows successful decoupling and feature extraction of dynamic and static information.

(3) Input the extracted dynamic grid features and local map information of static obstacles into the occupancy grid map prediction and generation module.

Step 3: The present invention uses a variational autoencoder (VAE) to build an OGM occupied grid map prediction generation module. The module obtains pre-trained network parameters through model training through extracted dynamic grid features and local maps, and predicts the future. A sequence of raster images at a moment in time, representing the distribution of the environment. Refer to Figures 2 and 3.

(1) Encoder: The encoder part consists of two convolutional layers and two residual layers. The first layer of convolution inputs features M with dimension 32 and outputs features with dimension 64. The convolution kernel size is 4, the step size is 2, and the padding value is 1. The second layer of convolution inputs features with a dimension of 64 and outputs features with a dimension of 128. The convolution kernel size is 4, the stride is 2, and the padding value is 1. The parameters of the two residual layers are consistent. Each residual layer inputs a 128-dimensional feature X. The feature X is processed by two layers of convolution to obtain X2. The output of the residual layer is Dimension, the output is 64 dimensions, the convolution kernel size is 3, and the stride is 1. The second layer of convolution inputs 64 dimensions and outputs 128 dimensions. The convolution kernel size is 1 and the stride is 1. Then the output of the residual layer is encoded through two convolution blocks with the same structure (input dimension 128, output dimension 2) to obtain μ and σ.

(2) Reshape μ and σ (dimension 16*16*2) into 512-dimensional vectors, and then generate distributions q(z|x):N(z|mu,z_var) and p(z): N(z|0,I), sample the distribution q(z|x) to obtain the latent vector Z in the variational autoencoder.

(3) Decoder: The decoder part uses the Z sampled from the normal distribution through the deconvolution module (input dimension is 2, output dimension is 128) to generate 128-dimensional features, and then inputs it into the VAE decoder. The decoder consists of two residual layers and two deconvolution layers. The residual layer structure is consistent with the structure in the encoder. It inputs 128-dimensional features and outputs 128-dimensional features. Then after the first convolution layer, the features of dimension 64 are output, the convolution kernel size is 4, and the stride is 2. Then the second convolution layer is input, and the output dimension is feature Q of 64, the convolution kernel size is 4, and the stride is 2.

(4) Q is input into the output convolution layer (convolution kernel size is 3, step size is 1), and the predicted occupancy grid sequence with dimension 1 is output and handed over to the motion planning module for processing.

Based on the above network structure, the present invention uses the entire training set to train the model, based on the KL divergence and cross-entropy hybrid loss function:

N is the number of samples, C is the number of categories, y _ij indicates whether the true label of the i-th sample is the j-th class (if it is 1, if not 0), p _ij indicates the prediction that the i-th sample belongs to the j-th class Probability. The present invention conducted 60 rounds of training, and each round traversed the entire training set. The gradient descent method is Adam, the batch_size is set to 64, and the learning rate is set to decrease with the increase of the number of training rounds. After 60 rounds of training, the loss function gradually decreases and stabilizes.

Step 4: After obtaining the predicted occupancy grid map sequence, the present invention uses each grid map sequentially as a feasible space for sampling candidate trajectories. A feasible spatial sequence is established that incorporates temporal information within the agent's local coordinate system. The present invention defines the planning problem as an optimization problem, which can be solved by obtaining the initial solution and iterative optimization. Please refer to Figure 4.

In order to simplify the calculation, the obstacle is expanded using the radius of the agent as the expansion radius, and the intelligent agent is converted into a particle. The purpose is to calculate the collision and determine the trajectory of the agent.

By obtaining the extended sequence of the occupied raster map, the occupied space in the predicted sequence can be derived through the position transformation relationship, which is shown in the following formula:

In a short time interval, the agent can be considered to be moving uniformly, the center of mass is the center of the circle, and [V _max Δt, V _min Δt] as the radius is the area that the agent can reach. Therefore, the feasible space representation of each frame can be expressed as follows: The states within the feasible interval must comply with strict speed interval constraints and collision constraints, and the final sampled trajectory must also remain within the feasible space. V _max and V _min represent the maximum and minimum speed of the agent respectively, Characterizes the spatial range that can be reached by driving at the maximum speed Δt time,/> It represents the spatial range that can be reached by driving at the minimum speed Δt time, and C _obs represents the space not occupied by obstacles.

Step 5: After obtaining the feasible space, generate candidate trajectories in it. The present invention uses a random sampling method to select frames within the feasible space. The present invention individually samples the expanded raster image sequence in step 4 one by one, and the sampling space represents the unoccupied area in each image. When sampling subsequent maps, the reachability of unoccupied grids is evaluated to ensure reachability. In order to satisfy the kinematic constraints of the agent, the search space is described as follows:

If the sampling points between adjacent frames can be connected without any collision, and the construction conditions are met: the connected edges do not fall outside the feasible space, and the dynamic constraints and kinematic constraints between the two points are met, then the requirements will be met Sample points between adjacent frames are connected together to form edges. If a sampling point does not meet these construction conditions, it is discarded, and then the list of sampling points is traversed to identify the node with the lowest cost and a backtracking process is initiated to obtain the initial trajectory as the initial solution.

Step 6: After obtaining the initial trajectory that satisfies both kinematic constraints and obstacle constraints, a smoother and safer trajectory is generated by using graph optimization methods. The agent's state transition at any given time satisfies the following conditions:

The evaluation functions to be optimized include lateral and longitudinal distance deviations, directional deviations, linear velocity deviations and angular velocity deviations from the reference trajectory. The optimization evaluation index L({P _t }) consists of two parts: the first part represents the deviation from the reference trajectory {P _ref } obtained by the global path, while the second part reflects the smoothness of the current path: L({P _t } )=L ^offset ({P _t }; {P _ref })+L ^smooth ({P _t }). where L({P _t }) represents the cost of the current path, L ^offset ({P _t }; {P _ref }) represents the offset from the reference trajectory {P _ref }, and L ^smooth ({P _t }) represents The smoothness of the trajectory.

Dynamic constraints can be expressed as ensuring that the agent's acceleration and velocity adhere to specific limits. Ultimately, the path planning problem is formulated as a nonlinear optimization problem:

Among them, L({P _t }|{m _t:t+β }) represents the optimization goal, {m _t:t+β } represents the grid map sequence, C _ego represents the outer contour of the self-car, and C _obs represents the outer contour of the obstacle. contour, represents the maximum speed and minimum speed that can be achieved, ω _min , ω _{max} represents the maximum angular velocity and minimum angular velocity that can be achieved, |a _v |≤a _{max} , which means that the linear acceleration is less than the maximum linear acceleration, Indicates that the angular acceleration is less than the maximum angular acceleration.

Taking into account the challenges associated with solving non-convex constrained problems, hard constraints are treated as soft constraints by converting them into penalties within the objective function. If the optimized trajectory cannot satisfy the kinematic constraints and obstacle constraints, the optimization process of this step is repeated for further optimization until the upper limit of the iteration is reached, then the solution process is ended and the final planned path is output.

The present invention is verified on a simulation platform based on Gazebo software. In order to correspond to multiple types of indoor environments, the present invention uses Gazebo software to build a simulation environment and build an indoor environment simulation platform including corridors, tables, and chairs. Each round of testing uses random starting and ending points, and movement speed. The starting position is randomly distributed with obstacles 0.5m away around the agent. The number of obstacles, maximum and minimum speeds can be set. When the distance between the obstacles and the agent If it is less than 0.05m, it will be regarded as a collision. If a collision occurs, the current agent test round will be deemed to be over. If the end point is reached normally, the current agent test round is deemed to be over.

Based on the above test requirements, the test parameters used in this invention are as shown in the following table:

The final test results are shown in the table below:

Based on the concept of the method of the present invention, a lightweight predictive navigation system suitable for indoor environments is also provided, including a pose prediction module, a dynamic and static feature extraction module, an occupancy grid sequence prediction module, and a feasible space acquisition module. Initial trajectory acquisition module and trajectory optimization module;

The pose prediction module performs coordinate conversion of sensory measurement values on the pose information, speed information and environmental representation of the intelligent body obtained by the multi-sensor system, and projects the environmental variables into the coordinate system of the future intelligent body;

The dynamic and static feature extraction module is used to consider the environmental variables in the coordinate system where the future agent is located, the propulsion of dynamic obstacles and the occupation of static obstacles in the environment, capture the interaction between the two, and combine the dynamic components with Decouple static components to extract latent variable features of future probability occupation raster maps;

The occupancy raster map sequence prediction module uses a variational autoencoder to build an OGM occupancy raster map prediction generation module, taking the latent variable characteristics of the future probability occupancy raster map as input, and the decoder outputs the predicted occupancy raster map sequence;

The feasible space acquisition module is used to sequentially use each raster map as a feasible space for sampling candidate trajectories;

The initial trajectory acquisition module is used to generate candidate trajectories in the feasible space and obtain initial trajectories that satisfy kinematic constraints and obstacle constraints;

The trajectory optimization module uses graph optimization methods to generate smoother and safer trajectories based on the initial trajectory.

On the other hand, the present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the lightweight indoor environment suitable for indoor environments described in the present invention can be realized. Predictive navigation methods.

The computer equipment may be a laptop computer, a desktop computer, a workstation or a vehicle-mounted computer.

The processor of the present invention may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or an off-the-shelf programmable gate array (FPGA).

The memory of the present invention can be an internal storage unit of a notebook computer, a desktop computer, a workstation or a vehicle-mounted computer, such as a memory or a hard disk; it can also be an external storage unit, such as a mobile hard disk or a flash memory card.

The present invention can also provide a computer device, including a processor and a memory. The memory is used to store computer executable programs. The processor reads the computer executable programs from the memory and executes them. When the processor executes the computer executable programs, it can Implement the lightweight predictive navigation method suitable for indoor environments according to the present invention.

Computer-readable storage media may include computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer-readable storage media may include: Read Only Memory (ROM), Random Access Memory (RAM), Solid State Drives (SSD), optical disks, etc. Among them, random access memory may include resistive random access memory (ReRAM, Resistance Random Access Memory) and dynamic random access memory (DRAM, Dynamic Random Access Memory).

To sum up, the present invention provides a lightweight predictive navigation method suitable for indoor environments. It adopts a hierarchical prediction-planning method to obtain sensing data and vehicle's own motion data (speed, coordinates) in a unified environment representation form. ) partially decouples the dynamic and static environments. The static environment obtains the future position through coordinate transformation and mapping. The dynamic environment predicts the raster map of several steps in the future by extracting scene motion feature information, and then directly uses the future raster to process the dynamic and static environments at the same time. Obstacles improve the efficiency and effectiveness of trajectory optimization.

The method of the present invention can perform long-term motion planning of an agent in a unified representation form, which is a difficult task. Raster map is a streamlined and common environment representation method, which has a representation similar to that in a top view and is suitable for unstructured environments. Raster maps can represent the feasible space required for planning, are abstract and highly reusable, and can be directly utilized for planning. The benefits of occupancy grid graphs are the ability to represent objects of arbitrary shapes and the ability to estimate the motion of objects without the need for explicit data associations. The method of the present invention uses a probabilistic occupancy raster diagram as a comprehensive representation to prevent information loss during the conversion process. The present invention chronologically organizes a sequence of predicted occupancy grid maps to generate trajectories, thereby alleviating the planning process' dependence on previously predicted information. Regarding prediction, by utilizing single-line lidar for sensing and combining motion information, the challenges of poor real-time performance and inaccurate prediction results are solved. In addition, a lightweight structure is used to speed up calculations. Regarding the planning problem, short-term planning problems are solved by generating feasible trajectories based on the occupancy grid sequence while taking into account the future distribution of obstacles. Furthermore, the obtained trajectories are optimized to ensure smoothness and reduce trajectory discontinuities.

Finally, it should be noted that the above is only to illustrate the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Those familiar with the technical field should understand that on the basis of the technical solution of the present invention, , modifications or deformations made based on the technical solution created by the present invention and its inventive concept shall be covered by the protection scope of the present invention.

Claims (10)

1. The lightweight predictive navigation method suitable for the indoor environment is characterized by comprising the following steps of:

the method comprises the steps of performing coordinate conversion of sensing measurement values on pose information, speed information and environmental characterization of an intelligent agent obtained by a multi-sensor system, and projecting environmental variables to a coordinate system where the intelligent agent is located in the future;

taking into consideration the self-walking of the dynamic barrier and the occupation of the static barrier in the environment of the environment variable under the coordinate system of the future intelligent agent, capturing the interaction between the dynamic barrier and the static barrier, decoupling the dynamic component and the static component, and extracting the potential variable characteristics of the future probability occupation grid map;

constructing an OGM occupation raster pattern prediction generation module by adopting a variation automatic encoder, taking potential variable characteristics of a future probability occupation raster pattern as input, and outputting a predicted occupation raster pattern sequence by a decoder;

sequentially using each raster pattern as a feasible space for sampling candidate trajectories;

generating candidate tracks in the feasible space, and acquiring initial tracks meeting kinematic constraint and barrier constraint;

a smoother and safer trajectory is generated based on the initial trajectory using a graph optimization method.

2. The lightweight predictive navigation method for an applicable indoor environment according to claim 1, wherein projecting the environmental variable into a coordinate system in which the future agent is located by performing coordinate transformation of the sensed measurement values of the agent pose information, the velocity information, and the environmental characterization obtained by the multi-sensor system comprises:

firstly, calculating the current obstacle occupation position according to the measured value of the laser radar, and acquiring the obstacle occupation position under the coordinate system of the current intelligent agent;

deducing the pose of the intelligent agent at the next moment according to the conversion relation between the current pose and the next moment of the intelligent agent and the current pose information of the intelligent agent; and projecting the current obstacle distribution to a coordinate system of the pose of the intelligent agent after the set time length, and obtaining the projection of the environment variable under the coordinate system of the intelligent agent in the future.

3. The lightweight predictive navigation method for an applicable indoor environment according to claim 1, wherein taking into account the occupancy of the environment by the self-moving and static obstacles of the dynamic obstacle, capturing the interaction between the two, and decoupling the dynamic and static components, extracting the potential variable features of the future probability occupancy trellis diagram comprises the steps of:

feeding the whole grid map information into a convolution long-short time memory network module, wherein the convolution long-short time memory network module takes a one-dimensional occupied grid graph as input, and generates dynamic grid features by using a convolution kernel with the size of 3;

predicting a local map of the static obstacle in the future by adopting a Bayesian generation method, and filtering out dynamic information to obtain local map information of the static obstacle;

and taking the extracted dynamic grid characteristics and the local map information of the static obstacle as the input of the OWM occupation grid map prediction generation module.

4. The lightweight predictive navigation method applicable to indoor environments according to claim 1, wherein the OGM occupying raster image prediction generation module comprises an encoder and a decoder, the encoder is composed of two convolution layers and two residual layers, and parameters of the two residual layers are consistent; the decoder consists of two residual layers and two deconvolution layers, the residual layer structure is identical to that in the encoder.

5. The lightweight predictive navigation method for an applicable indoor environment according to claim 1, wherein when each raster pattern is sequentially used as a feasible space for sampling candidate trajectories, specifically, based on an extended sequence of occupied raster patterns, the occupied space in the predicted sequence is derived through a position transformation relationship, in a set period of time, an agent uniformly moves, a centroid is a center of a circle, a feasible space representation of each frame is obtained through an area that the agent can reach, and states in the feasible interval obey a strict speed interval constraint and a collision constraint.

6. The lightweight predictive navigation method applicable to indoor environments according to claim 1, wherein candidate tracks are generated in the feasible space, frames are selected in the feasible space by adopting a random sampling method when initial tracks meeting kinematic constraint and barrier constraint are obtained, grid graph sequences are sampled individually and one by one, the sampled space represents an unoccupied area in each graph, and a search space is obtained on the basis of meeting the kinematic constraint of an intelligent agent and ensuring accessibility during sampling; when sampling is carried out, if sampling points between adjacent frames can be connected under the condition of no collision, judging whether the connected edges meet the dynamic constraint and the kinematic constraint of an intelligent agent, and if so, connecting the sampling points together to form the edges; traversing the node list to identify the lowest cost node and initiating a backtracking process to obtain an initial trajectory that satisfies the kinematic and obstacle constraints.

7. The lightweight predictive navigation method for an applicable indoor environment of claim 1, wherein generating a smoother, safer trajectory based on the initial trajectory using a graph optimization method comprises:

creating smoother and safer trajectories using the graph optimization method is formulated as a nonlinear optimization problem:

wherein L ({ P) _t }|{m _t:t+β }) represents optimization objective, { m _t:t+β "represents a raster pattern sequence, C _ego Represents the outline of the bicycle, C _obs Represents the outline of the obstacle and,represents the maximum and minimum speeds achievable, ω _min ,ω _{max} Represents the maximum and minimum angular velocities achievable, |a _v |≤a _{max} Indicating that the linear acceleration is less than the maximum linear acceleration, indicating that the angular acceleration is less than the maximum angular acceleration;

converting the hard constraint into punishment in the objective function, and treating the hard constraint as a soft constraint; and (3) performing iterative optimization until the optimized track meets kinematic constraint and barrier constraint or reaches the iterative upper limit, and calculating to obtain a final planning path.

8. The lightweight predictive navigation system suitable for the indoor environment is characterized by comprising a pose prediction module, a dynamic and static feature extraction module, an occupied grid pattern sequence prediction module, a feasible space acquisition module, an initial track acquisition module and a track optimization module;

the pose prediction module is used for carrying out coordinate conversion on sensing measurement values on pose information, speed information and environment characterization of the intelligent agent obtained by the multi-sensor system, and projecting the environment variables to a coordinate system where the intelligent agent is located in the future;

the dynamic and static feature extraction module is used for taking into account the occupation of the dynamic barrier and the static barrier in the environment by the self-walking of the environment variable under the coordinate system of the future intelligent agent, capturing the interaction between the dynamic component and the static component, decoupling the dynamic component and the static component, and extracting the potential variable feature of the future probability occupation grid map;

the occupied raster pattern sequence prediction module adopts a variation automatic encoder to construct an OWM occupied raster pattern prediction generation module, potential variable characteristics of future probability occupied raster patterns are used as input, and a decoder outputs a predicted occupied raster pattern sequence;

the feasible space acquisition module is used for sequentially using each grid graph as a feasible space of the sampling candidate track;

the initial track acquisition module is used for generating candidate tracks in the feasible space and acquiring initial tracks meeting kinematic constraint and obstacle constraint;

the track optimization module adopts a graph optimization method to generate smoother and safer tracks based on the initial tracks.

9. A computer device comprising a processor and a memory, the memory storing a computer executable program, the processor reading the computer executable program from the memory and executing the program, the processor executing the program to implement the lightweight predictive navigation method of any one of claims 1-7 for use in an indoor environment.

10. A computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, the computer program, when executed by a processor, is capable of implementing the lightweight predictive navigation method of any one of claims 1-7, which is applicable to indoor environments.