Logistics unmanned vehicle express fast arrival method in park scene
By using 3D passion radar and improved YOLOv5 model and TD3 reinforced learning algorithm in logistics unmanned vehicles, the autonomous navigation and traffic sign detection of park logistics unmanned vehicles are realized, solving the inefficiency and safety risks of traditional logistics transportation methods. The problem has improved the adaptability and reliability of logistics and transportation in the park.
Patent Information
- Application Number
- CN202410738457.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-07
- Publication Date
- 2024-11-26
AI Technical Summary
In the park environment, traditional logistics transportation methods face problems such as high human resources costs, low efficiency and traffic congestion, and the success rate of existing logistics unmanned vehicle systems to comply with traffic rules and multi-task execution is not high when moving at high speeds.
Using 3D sexual radar sensor to obtain 3D laser point cloud information, combined with global path planning algorithm and reinforcement learning technology, realize autonomous navigation and traffic marking detection of logistics unmanned vehicles, and generate local paths through the improved YOLOv5 model and TD3 reinforcement learning algorithm. Tracking, ensuring that logistics unmanned vehicles comply with traffic rules and perform multitasking when moving at high speeds.
It improves the adaptability and reliability of logistics unmanned vehicles in the park environment, ensures the compliance with traffic rules and the success rate of multi-task execution during high-speed movement, and improves transportation efficiency and safety.
Smart Images

Figure CN119026765A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to autonomous navigation, and more particularly to a method for express delivery of logistics unmanned vehicles in a park scenario. Background Art
[0002] In the park environment, the demand for logistics and transportation is growing. However, traditional logistics and transportation methods face many challenges, including high human resource costs, low efficiency, and traffic congestion. Therefore, the automated logistics and transportation system has become one of the keys to solving these problems.
[0003] In the past, campus logistics transportation usually relied on manual driving or fixed-route automated navigation systems. However, these traditional methods have certain limitations. For example, manual driving is easily affected by human factors, resulting in low transportation efficiency and safety hazards; while fixed-route automated navigation systems lack flexibility and cannot adapt to dynamic changes in the environment.
[0004] Therefore, in order to improve the efficiency and safety of park logistics and transportation, researchers and enterprises have been exploring new technologies and methods. In recent years, the development of unmanned vehicle technology has brought new solutions to park logistics and transportation. Unmanned vehicles are equipped with various sensors and intelligent algorithms to achieve functions such as autonomous navigation, obstacle avoidance and traffic rules compliance, which greatly improves the efficiency and safety of logistics and transportation.
[0005] However, there are still some problems with the current unmanned logistics vehicle systems on the market, such as low compliance with traffic rules and low success rate of multi-task execution under high-speed movement. Therefore, how to design an unmanned logistics vehicle system that integrates mapping, perception and planning to improve its adaptability and reliability in the park environment has become one of the current research focuses. Summary of the invention
[0006] The purpose of the present invention is to propose a method for express delivery of logistics unmanned vehicles in a park scenario.
[0007] The technical solution to achieve the purpose of the present invention is: a method for express delivery of unmanned logistics vehicles in a park scenario, the steps are as follows:
[0008] Step 1: Use the 3D laser radar sensor to obtain 3D laser point cloud information, build a laser point cloud map M of the park scene, and obtain the real-time position information P of the logistics unmanned vehicle. c ;
[0009] Step 2: Based on the constructed laser point cloud map M and the global target point G, a global path planning algorithm hybridA* is used to generate a global trajectory P g ;
[0010] Step 3: Based on the constructed laser point cloud map M, combined with the real-time pose information P c , calling the YOLOv5 model to detect traffic signs T on the image I of the park scene during driving;
[0011] Step 4: Based on the real-time detected traffic sign T, combined with the specific traffic rules of the road and the instructions for picking up and delivering goods, publish the local target point W l , change the instant driving strategy S, adopt the local path planning based on TD3 reinforcement learning, and generate a new local path trajectory P l ;
[0012] Step 5: Execute the local path P l Afterwards, execute the global path P g , and continue to collect perception information until reaching the end point G.
[0013] Further, in step 1, the 3D laser point cloud information is obtained by using the 3D laser radar sensor, a laser point cloud map M of the park scene is established, and the real-time position information P of the unmanned vehicle is obtained. c , the specific method is:
[0014] The 3D laser point cloud information is obtained by using the 3D laser radar sensor, and the LeGO-LOAM algorithm is used to build the laser point cloud map M of the park scene. The laser point cloud map P contains static obstacles O. s and traffic signs T; the ndt matching positioning algorithm is used to match the previously established laser point cloud map M to obtain the current location information P of the logistics unmanned vehicle c .
[0015] Further, step 3, based on the constructed laser point cloud map M, combined with the real-time pose information P c , calling the YOLOv5 model, detecting the traffic sign T on the image I of the park scene during driving, extracting the traffic sign features h under the park driving rules, and providing perception information for the express delivery task of the logistics unmanned vehicle. The specific method is as follows:
[0016] The YOLOv5 model consists of three parts: Backbone, Neck, and detection head. Backbone uses the CSPDarkNet53 network to extract features from the real-time image I collected by the monocular camera during the driving of the logistics unmanned vehicle to obtain multi-scale feature information n 1 ; Neck uses PAFPN to fuse multi-scale features of real-time images extracted from Backbone 1 ; The detection head outputs multi-scale features n according to Neck 1Predict the target category c and set the prediction box b. Backbone replaces the C3 module with the CSPCM module to achieve the purpose of optimizing the Backbone network structure. At the same time, the CA attention mechanism for specific targets is introduced to focus attention on the target area R. After detecting the traffic sign T, the detection area is solidified and the weight of the key feature points is increased to enhance the model's recognition accuracy of the traffic sign T.
[0017] (1) Backbone replaces the C3 module with the CSPCM module to achieve the purpose of optimizing the Backbone network structure;
[0018] The input image I of the logistics unmanned vehicle is first sent to CSPDarkNet53 for preliminary multi-scale feature extraction, and then these multi-scale features n 1 The feature information is transferred to the CSPCM module, which enhances the feature expression capability by stacking ConvMixer layers and 1x1 convolution layers, and performs feature fusion and splicing in the channel dimension. 2 It is then input into the Neck part for multi-scale feature fusion and then sent to the detection head for accurate prediction of the target category c and generation of the prediction box b.
[0019] (2) Introducing a CA attention mechanism for specific targets to focus attention on the target area R. After detecting the traffic sign T, the detection area is solidified and the weights of key feature points are increased, thereby enhancing the recognition accuracy of the model for the traffic sign T.
[0020] The feature map output by CSPDarkNet53 is used as the input of the CA attention module. Then, the global statistical information in the feature map is extracted through global average pooling, and the fully connected layer is used to transform and reduce the dimension of the pooled global statistical information. Subsequently, nonlinearity is introduced through the activation function to further activate and refine the feature representation. Finally, the key features of the target area R are highlighted through feature enhancement, while the feature response of the non-target area is suppressed to generate a weighted feature map and send it to the Neck part.
[0021] Among them, the key features of the target area R are highlighted by feature enhancement, while the feature response of the non-target area is suppressed to generate a weighted feature map. Specifically, according to the relative distance information between the logistics unmanned vehicle and the target, the weight of the feature point is dynamically adjusted, and the close feature points are given higher weights, while the weights of the long-distance feature points are gradually reduced. The hyperbolic tangent function used by the weight value function is as follows:
[0022] p = tanh(d)
[0023] Among them, d is the relative distance between the logistics unmanned vehicle and the target, and p is the weight value of each feature point.
[0024] Furthermore, before using the YOLOv5 model to detect the target in the input image or video frame, the target detection frame correction strategy is adopted to perform geometric correction on the detection frame. The corrected target detection frame makes the static feature points appear on the corresponding optical flow epipolar line l1, while the dynamic target has a large deviation from the epipolar line l1. The specific method is as follows:
[0025] The target detection frame correction strategy comprehensively collects the posture information, position and speed information, and the distance information between the logistics unmanned vehicle and the target point, constructs the posture state conversion matrix, and corrects the target bounding box. The target detection frame geometry correction function is as follows:
[0026] x′=f(Δx)+x
[0027] y′=f(Δy)+y
[0028] w′=p w ·(Δw)·w
[0029] h′=p h ·(Δh)·h
[0030] Among them, the coordinates of the center point of the original detection frame are (x, y), the width and height are w and h respectively, and the offsets output by the YOLOv5 model are Δx, Δy, Δw, Δh. Through the target detection frame geometry correction function, the coordinate values of the corrected target detection frame are:
[0031] x′ 1 =T{11}·x 1 +T{12}·y 1 +T{13}
[0032] y′ 1 =T{21}·x 1 +T{22}·y 1 +T{23}
[0033] x′ 2 =T{11}·x 2 +T{12}·y 2 +T{13}
[0034] y′ 2 =T{21}·x 2 +T{22}·y 2 +T{23}
[0035] x′ 3 =T{11}·x 3 +T{12}·y 3 +T{13}
[0036] y′ 3 =T{21}·x3 +T{22}·y 3 +T{23}
[0037] x′ 4 =T{11}·x 4 +T{12}·y 4 +T{13}
[0038] y′ 4 =T{21}·x 4 +T{22}·y 4 +T{23}
[0039] Among them, (x1, y1), (x2, y2), (x3, y3), (x4, y4) are the coordinate values of the four vertices of the detection box, T{n 1 n 2} is point n 1 To point n 2 On this basis, the YOLOv5 model is used to detect the target area and distinguish the background area and traffic sign T.
[0040] Furthermore, the YOLOv5 model is used to perform target detection on the input image or video frame. After obtaining a series of initially recognized traffic sign T bounding boxes and corresponding confidence scores, a pseudo target point removal algorithm is executed. The specific method is as follows:
[0041] Use the YOLOv5 model to detect the input image or video frame, obtain a series of initially identified traffic sign T bounding boxes and corresponding confidence scores, and then execute the pseudo target point elimination algorithm. Use the optical flow estimation technology to track the motion trajectory of the target area feature points, that is, the traffic sign T, analyze the kinematic model of the target area feature points in continuous frames, establish a dynamic motion model of each feature point, compare it with the preset threshold, identify abnormal points that do not conform to the kinematic characteristics of the traffic sign T, and mark them as potential pseudo target points. Calculate the relative average speed of the traffic sign T relative to the vehicle body, and construct the kinematic model as follows:
[0042]
[0043] Among them, E x 、E y The components of the relative average velocity of the target area relative to the vehicle body along the x and y directions, N is the number of feature points for calculating the pose information in the target area, and u j 、v j Calculate the components of the feature point motion velocity along the x and y directions for the vehicle body relative to the posture information.
[0044] Further, in step 4, according to the real-time detected traffic sign T, combined with traffic rules and cargo pickup and delivery instructions, a local target point W is published. l , change the instant driving strategy S, adopt the local path planning based on reinforcement learning, and generate a new local path trajectory P l , the specific method is:
[0045] Logistics unmanned vehicle environment information obtains local environmental information through laser radar e , and the polar coordinate information of the target relative to the logistics unmanned vehicle in polar coordinates Combination, where d represents the Euclidean distance between the current position of the unmanned vehicle and the target position, Represents the interpolation angle between the current angle of the unmanned vehicle and the target position, s e and r The combined data will be used as input state s in the action network of TD3;
[0046] The state s is input into the action network of the reinforcement learning network. The action network consists of two fully connected layers (FC). After each fully connected layer, it is activated by ReLU (activation function). The activation layer is connected to the output layer, and the two action parameters are output, which are the linear speed a of the logistics unmanned vehicle and 1 and angular velocity a 2 ; Then, the output value of the output layer is controlled to (-1,1) through tanh (hyperbolic tangent function), and then action a is applied to the logistics unmanned vehicle. The specific formula is:
[0047]
[0048] Among them, v max is the maximum linear velocity, ω max is the maximum angular velocity;
[0049] After the action network obtains the action of the unmanned vehicle, the Q value of the action network is evaluated by two evaluation networks. The two evaluation networks have the same structure. The state s and action a are used as the input of the evaluation network. The state s is input to the fully connected layer of the evaluation network. The fully connected layer is activated by ReLU and outputs L s Output L s and actions will be input into independent transformation fully connected layers τ with the same size 1 and τ 2 The combination is as follows:
[0050]
[0051] Among them, L C is the output of the combined fully connected layer (CFC), and They are the fully connected layers τ1 and τ 2 The weight of is the fully connected layer τ 2 The bias of ; ReLU is applied to the combination layer and 1 parameter is output to represent the Q value; the smallest Q value of the two evaluation networks is selected as the final evaluation parameter output to limit the overestimation of the Q value by the actor-critic network;
[0052] The reward function is as follows:
[0053]
[0054] The state-action pair (s t ,a t ) depends on three conditions. If at the current time step D t The distance to the target is less than the threshold η D , then a positive reward r is used g ; If a collision is detected, a negative reward r is used c ; If none of them are satisfied, an immediate reward is used based on the current linear velocity v and angular velocity ω;
[0055] In order to guide the navigation strategy towards the set goal and obtain different parameter values, the parameters of the two are updated delayedly. The delayed attribute reward method is used to obtain a more stable reward value. The specific formula is as follows:
[0056]
[0057] Among them, n is the number of previous steps to update the reward. The positive reward is not only determined by the state-action when the goal is reached, but also decreases in the last n steps before reaching the goal. Through the constraint of the reward function on the network parameters, the reinforcement learning network learns that the logistics unmanned vehicle moves from the current position to the local target point W. l The local path trajectory is merged and updated into the final global trajectory until the global target point G is reached.
[0058] A logistics unmanned vehicle express delivery system in a park scenario implements the logistics unmanned vehicle express delivery method in a park scenario to achieve logistics unmanned vehicle express delivery in a park scenario.
[0059] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for express delivery of unmanned logistics vehicles in a park scenario is implemented to achieve express delivery of unmanned logistics vehicles in a park scenario.
[0060] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the method for express delivery of unmanned logistics vehicles in a park scenario is implemented to achieve express delivery of unmanned logistics vehicles in a park scenario.
[0061] Compared with the prior art, the present invention has the following significant advantages: 1) In view of the fact that logistics unmanned vehicles in park environments have limited computing resources and need to comply with traffic rules when moving at high speeds, and the success rate of performing multiple tasks is not high, a set of logistics unmanned vehicle pickup and delivery solutions that integrates mapping, perception, and planning is designed. 2) Considering the perspective and background changes of targets in high-speed dynamic environments, a CA attention mechanism for specific targets and an improved pseudo-target elimination algorithm are designed to ensure the accuracy of the detection network. 3) In order to ensure the stable execution of multiple tasks, a global path planner based on the Hybrid A* algorithm and a local path planner based on the TD3 reinforcement learning algorithm are designed. By presetting traffic rules and combining real-time perception and positioning information, autonomous navigation and task execution of logistics unmanned vehicles in logistics parks are realized. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 It is a flowchart of the method for express delivery of unmanned logistics vehicles in a park scenario of the present invention;
[0063] Figure 2 It is a flow chart of the target detection thread method of the present invention; DETAILED DESCRIPTION
[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0065] The present invention provides a logistics unmanned vehicle express delivery method in a park scenario, comprising the following steps:
[0066] Step 1: The logistics unmanned vehicle uses the open source laser mapping and positioning algorithm LeGO-LOAM to locate and map the 3D laser point cloud information obtained by the 3D laser radar sensor, build a laser point cloud map M of the park scene, and obtain the real-time position information P of the unmanned vehicle. c , providing positioning guarantee for the unmanned vehicle's subsequent express delivery mission.
[0067] Step 2: The logistics unmanned vehicle generates a global trajectory P based on the pre-built map M and the global target point C using the global path planning algorithm hybrid A* algorithm g .
[0068] The global path planning module generates a global path based on the previously established laser point cloud map P, which contains static obstacles O s And traffic signs T. In the map P, find a global trajectory P from the starting point to the global target point G for a logistics unmanned vehicle such as a mobile robot or vehicle. g 。 Hybrid A * The algorithm combines the traditional A * Search algorithms and continuous state space search methods can effectively handle problems involving continuous state spaces.
[0069] Furthermore, Hybrid A * The algorithm uses A * The heuristic function and cost estimate of the search algorithm are used to search the discretized map, gradually expanding from the starting point to the target point to find the shortest path. The specific function is as follows:
[0070] f(n)=g(n)+h(n)
[0071] Where f(n) is the estimated shortest path cost from the start node to the target node through node n. g(n) is the actual cost from the start node to node n, that is, the length of the path that has been traveled. h(n) is the heuristic estimated cost from node n to the target node, that is, the estimate of the remaining path length.
[0072] In A * Based on the search, Hybrid A * The algorithm introduces continuous state space search, optimizes the path, performs sampling in the continuous state space, and generates new candidate paths. It also determines whether the path is feasible based on the collision between the lines between the sampling points and the map obstacles. * Among all the generated paths, the path with the lowest cost is selected as the current optimal path. After an optimal path is determined, Hybrid A * Path extraction will be performed. Through the steps of path smoothing, path connection, path sampling and path length adjustment, the continuous optimal path is converted into a series of discrete path points or global trajectory P g , providing a practical and executable navigation route for logistics unmanned vehicles.
[0073] Step 3: The logistics unmanned vehicle uses the pre-built map M and real-time posture information P c , calling the pre-trained improved YOLOv5 algorithm, performs real-time detection of the image I of the park scene during driving, detects traffic signs T, extracts traffic sign features h under the park driving rules, and provides perception information for the express delivery mission of the logistics unmanned vehicle.
[0074] The YOLOv5 network structure is mainly composed of three parts: Backbone, Neck and detection head. Backbone uses the CSPDarkNet53 network to extract features from the real-time image I collected by the monocular camera during the driving of the logistics unmanned vehicle to obtain multi-scale feature information n 1 ; Neck uses PAFPN to fuse multi-scale features of real-time images extracted from Backbone 1 ; The multi-scale feature n output by the detection head to Neck 1 Predict the target category c and set the prediction box b. However, when the logistics unmanned vehicle is driving, static targets such as traffic signs may be identified as dynamic targets d, thereby affecting the feature extraction of static targets s and further affecting the execution of subsequent tasks.
[0075] Therefore, the present invention achieves the purpose of optimizing the Backbone network structure by replacing the original C3 module in YOLOv5 with the CSPCM module. The input image I of the logistics unmanned vehicle is first sent to CSPDarkNet53 for preliminary feature extraction, and then these multi-scale features n 1 The information is passed to the CSPCM module, which enhances the expressiveness of features by stacking ConvMixer layers and 1x1 convolution layers, and performs feature fusion and concatenation in the channel dimension. 2 Then it is input into the Neck part to realize the multi-scale feature n 1 Finally, these optimized features n 2 It is sent to the detection head for accurate prediction of the target category c and generation of the prediction box b, thereby realizing accurate detection and recognition of traffic signs T in the park scene in the unmanned vehicle express delivery mission. Through the improvement strategy proposed in this patent, the perception ability and target recognition accuracy of the unmanned vehicle in a dynamic environment have been significantly improved, providing strong technical support for the efficient execution of the express delivery mission.
[0076] At the same time, in order to solve the unfavorable situation that the feature points gradually decrease as the traffic sign T approaches, a CA attention mechanism for specific targets is introduced. This mechanism aims to focus the model's attention on the target area R, solidify the detection area after detecting the traffic sign T, and increase the weight of the key feature points, thereby enhancing the model's recognition accuracy of the traffic sign T. In order to introduce the CA attention mechanism into YOLOv5, it is necessary to determine the integration point first. This project determines the integration point p after the convolution layer and before the activation function.
[0077] Specifically, the integration point p of the CA attention mechanism is precisely located in the Backbone network structure in the YOLOv5 model, that is, after the output of CSPDarkNet53 and before the Neck structure. At this integration point, the feature map output by CSPDarkNet53 is used as the input of the CA attention module. The feature map contains rich channel information and spatial information. Then, the global statistical information in the feature map is extracted by global average pooling, and the pooled features are transformed and reduced in dimension using the fully connected layer. Subsequently, nonlinearity is introduced through the activation function to further activate and refine the feature representation. Finally, the feature enhancement step highlights the key features of the target area R while suppressing the feature response of the non-target area to generate a weighted feature map, which is then sent to the Neck structure of the model for subsequent multi-scale feature fusion and target prediction, thereby improving the model's detection performance for traffic signs T, especially maintaining high-accuracy detection capabilities when the target is approaching.
[0078] After the CA attention mechanism is introduced, the hyperparameters of YOLOv5 need to be adjusted to obtain the best training effect. The modified YOLOv5 model is trained using the training dataset, and the performance of the model is evaluated using the validation dataset. Based on the validation results, the CA attention module or other parts of YOLOv5 are further optimized.
[0079] When YOLOv5 with CA attention mechanism achieves the ideal accuracy on the validation set, the final test is conducted on the test set, and necessary adjustments are made according to the test results. After the test is successful, the model can be deployed to identify the traffic sign T in the logistics park. The improved YOLOv5 algorithm is deployed on the logistics unmanned vehicle, and the performance of the optimized model is monitored in the logistics park to ensure that it meets the expected reasoning speed and resource consumption requirements.
[0080] The present invention also proposes an improved pseudo target rejection algorithm, which specifically corrects the geometric changes of the target under different postures to effectively solve the problems of false detection and rejection failure of dynamic targets. The algorithm is based on the target detection frame correction strategy with dynamic posture changes, specifically including a pseudo target point rejection algorithm and a target detection frame correction strategy, aiming to improve the accuracy and robustness of logistics unmanned vehicles in detecting traffic signs T in park logistics scenarios.
[0081] When the logistics unmanned vehicle performs express delivery tasks within the park, it first uses the YOLOv5 model to perform target detection on the input image or video frame to obtain a series of preliminary recognized traffic sign T bounding boxes and corresponding confidence scores.
[0082] Based on this, the algorithm conducts refined processing and deeply analyzes the position changes of the traffic sign T detected by YOLOv5 in consecutive frames. It tracks the motion trajectory of the feature points in the target area through optical flow estimation technology, identifies the motion characteristics of the traffic sign T, and eliminates pseudo-target points caused by noise, occlusion or other interference factors. This dynamic analysis process is the key to identifying and distinguishing real targets from pseudo-targets.
[0083] The design of the pseudo target point elimination algorithm is based on a deep understanding of the target motion characteristics. The algorithm takes into account the possible motion patterns of traffic signs T under normal use conditions, and establishes a preset threshold and motion model accordingly. When the logistics unmanned vehicle performs tasks, the algorithm analyzes the motion patterns of the feature points in the target area in continuous frames and establishes a dynamic motion model. By comparing with the preset threshold, abnormal points that do not conform to the typical motion characteristics of traffic signs T are identified and marked as potential pseudo target points.
[0084] On the basis of the pseudo target point elimination algorithm, a target detection frame correction strategy is added to further improve the accuracy of traffic sign T recognition. The target detection frame correction strategy comprehensively collects the posture information, position and speed information, and the distance information between the logistics unmanned vehicle and the target point to construct a transformation matrix. This transformation matrix is used as input to correct the target bounding box detected by YOLOv5. The designed target detection frame geometry correction function is used to achieve precise adjustment of the target bounding box. The designed detection frame geometry correction function is as follows:
[0085] x′=f(Δx)+x
[0086] y′=f(Δy)+y
[0087] w′=p w ·(Δw)·w
[0088] h′=p h ·(Δh)·h
[0089] The coordinates of the center point of the original detection frame are (x, y), and the width and height are w and h respectively. The offsets output by YOLOv5 are Δx, Δy, Δw, Δh. This function can calculate the coordinate values of the corrected target detection frame in real time.
[0090] x′ 1 =T{11}·x 1 +T{12}·y 1 +T{13}
[0091] y′ 1 =T{21}·x 1 +T{22}·y 1 +T{23}
[0092] x′ 2 =T{11}·x 2 +T{12}·y 2 +T{13}
[0093] y′ 2 =T{21}·x 2 +T{22}·y 2 +T{23}
[0094] x′ 3 =T{11}·x 3 +T{12}·y 3 +T{13}
[0095] y′ 3 =T{21}·x 3 +T{22}·y 3 +T{23}
[0096] x′ 4 =T{11}·x 4 +T{12}·y 4 +T{13}
[0097] y′ 4 =T{21}·x 4 +T{22}·y 4 +T{23}
[0098] Among them, (x1, y1), (x2, y2), (x3, y3), (x4, y4) are the coordinate values of the four vertices of the detection frame respectively, and T is the transformation matrix calculated according to the collected posture information, target distance information and other parameters, which is used to correct the position of the original detection frame.
[0099] The geometrically corrected target detection frame makes the static feature points appear on the corresponding optical flow epipolar line l1. The dynamic target is judged as a dynamic point and is removed because of its large deviation from the epipolar line l1.
[0100] On this basis, the optimized YOLOv5 preprocessing model is used to detect typical target areas and distinguish between background areas and traffic signs T.
[0101] Calculate the relative average speed of the traffic sign T to the vehicle body:
[0102]
[0103] Among them, E x 、E y The components of the relative average velocity of the target area relative to the vehicle body along the x and y directions, N is the number of feature points for calculating the pose information in the target area, and uj 、v j Calculate the components of the feature point motion velocity along the x and y directions for the vehicle body relative to the posture information.
[0104] In addition, the optimized YOLOv5 preprocessing model is used to accurately distinguish traffic signs T from background areas, providing the logistics unmanned vehicle with a clear traffic sign T area. The logistics unmanned vehicle dynamically adjusts the weight of feature points based on the relative distance information to the target, giving higher weights to close feature points and gradually reducing the weights of long-distance feature points.
[0105] The hyperbolic tangent function used by the hierarchical weight value function is as follows:
[0106] p = tanh(d)
[0107] Among them, d is the relative distance between the logistics unmanned vehicle and the target, and p is the weight value of each feature point.
[0108] Finally, the traffic sign T detection results after pseudo target point elimination and weight optimization are weighted and integrated to provide accurate traffic sign T information for the logistics unmanned vehicle. This information is crucial for the navigation decision, path planning and traffic sign recognition functions of the logistics unmanned vehicle. Through continuous algorithm optimization and iteration, the robustness of the algorithm in different environments and conditions is ensured, and the safety and efficiency of the logistics unmanned vehicle in practical applications are improved.
[0109] Step 4: When the unmanned vehicle is driving, it publishes the local target point W according to the traffic sign T detected in real time, combined with traffic rules and cargo delivery instructions. l , change the instant driving strategy S, adopt the local path planning based on reinforcement learning, and generate a new local path trajectory P l .
[0110] First, the task instruction and target recognition module is called. The unmanned vehicle receives the task instruction from the task management system, including the geographical location of the default pickup point and the default delivery point, the task content, and other information. Then, based on this information, the Hybrid A* global path planning algorithm is called to calculate the initial global route P g , and use the ndt matching positioning algorithm to match the previously established laser point cloud map M to obtain the current position information P of the unmanned vehicle c The module output is the unmanned vehicle's driving strategy S and the local target point W. l , as the basis for the execution of subsequent tasks.
[0111] During the driving process, the unmanned vehicle synchronously calls the target detection module to obtain information about the road and surrounding environment and the switching information of the loading and unloading points through the camera and lidar. The camera analyzes and identifies the traffic signs T on the road through the improved target detection YOLOv5 algorithm, and switches the operation instructions of the loading and unloading points in real time. The lidar obtains the distance information d by measuring the unmanned vehicle and the traffic sign T, which provides a perception basis for the subsequent global and local path adjustments of the unmanned vehicle. The module output results are the traffic sign feature h and the local target point W corrected by the loading and unloading point information. l , as the input for the formulation of driving strategy S.
[0112] Finally, the driving strategy formulation and optimal path planning module is executed, combining the previously obtained traffic sign feature h and the local target point W corrected by the loading and unloading point information l Formulate driving strategy S and change the local point W of the unmanned vehicle l By adopting the local path planner based on reinforcement learning, the local path trajectory P is replanned l , ensuring that the logistics unmanned vehicle adopts the correct driving strategy in different situations, such as traffic lights on the road. l Afterwards, rerun the initial global route P g , update to get the current optimal global route P g , and collects perception information in real time to continuously update the optimal global path P g , until reaching the end point G.
[0113] The path generated by global path planning is based on the static prior LiDAR point cloud map M. However, during the driving process of the logistics unmanned vehicle, the driving strategy will change and the local target point W will be updated in real time. l At this time, the local path planner is needed to generate a new local path trajectory P l .
[0114] In the planning-based navigation stack, local path planning executes the commands of the local planner. This patent uses the TD3 reinforcement learning algorithm to train the local path planner. TD3 is a neural network architecture based on double-delayed deep deterministic policy gradient. It is an actor-critic network. The network is divided into two parts: the action network and the evaluation network. At the same time, the action and evaluation networks each contain two networks. The network allows continuous movement in space. The network learns a local navigation strategy and can reach the local target point W. l At the same time, it can also achieve obstacle avoidance through laser. First, the motion strategy is trained in a simulated environment using a TD3-based neural network architecture. The specific process is as follows:
[0115] The logistics unmanned vehicle obtains information about the local environment through lidar e , in polar coordinates, the polar coordinate information of the waypoint relative to the logistics unmanned vehicle Combination. Among them, d represents the Euclidean distance between the current position of the unmanned vehicle and the target position. Represents the angle interpolation between the current angle of the unmanned vehicle and the target position. The resulting combined data will be used as the input state s in the action network of TD3.
[0116] The state s is input into the action network of the reinforcement learning network. The action network consists of two fully connected layers (FC). After each fully connected layer, it will be activated by ReLU (activation function). The activation layer is connected to the output layer, and the two action parameters are output respectively as linear speed a and 1 and angular velocity a 2 .
[0117] Next, the output value of the output layer is controlled to (-1,1) through tanh (hyperbolic tangent function). Before applying the action to the environment, the specific formula is:
[0118]
[0119] Among them, v max is the maximum linear velocity, ω max is the maximum angular velocity.
[0120] After the action network obtains the action of the unmanned vehicle, the Q value of the action network is evaluated by two evaluation networks. The two evaluation networks have the same structure, but in order to obtain different parameter values, the parameters of the two networks will be updated delayed.
[0121] The state s and action a are used as the input of the evaluation network. The state s is input to the fully connected layer of the evaluation network, which outputs L after being activated by ReLU. s .
[0122] The output and action of the fully connected layer will be input to τ with the same size respectively. 1 and τ 2 The independent transformation fully connected layer (TFC) of is combined into the following formula:
[0123]
[0124] Among them, L C is a combined fully connected layer (CFC), and They are τ 1 and τ 2 The weight of . is τ 2 deviation.
[0125] ReLU is applied to the combined layer and outputs 1 parameter representing the Q value.
[0126] The smallest Q-value of the two evaluation networks is selected as the final evaluation parameter output to limit the excessive overestimation of the Q-value by the actor-critic network.
[0127] The reward function of this strategy is as follows:
[0128]
[0129] The state-action pair (s t ,a t ) depends on three conditions. If at the current time step D t The distance to the target is less than the threshold η D , then a positive reward r is used g If a collision is detected, a negative reward r is used c If neither is satisfied, an immediate reward is used based on the current linear velocity v and angular velocity ω.
[0130] In order to guide the navigation strategy towards the set goal and obtain different parameter values, the parameters of the two are updated delayedly. The delayed attribute reward method is used to obtain a more stable reward value. The specific formula is as follows:
[0131]
[0132] Where n is the number of previous steps to update the reward. The positive reward is not only determined by the state-action when the goal is reached, but also decreases on the last n steps before reaching the goal. By constraining the network parameters with the reward function, the reinforcement learning network can successfully learn the logistics unmanned vehicle from the current position to the local target point W. l The local path trajectory is merged and updated into the final global trajectory until the global target point G is reached.
[0133] Step 5: After executing the local path P l After that, execute the global path P in step 3 g , and continue to collect perception information until reaching the end point G.
[0134] In summary, the present invention uses the laser sensor simultaneous positioning and mapping (SLAM) technology based on the 3D multi-line laser radar carried by the unmanned vehicle to construct a point cloud map of the park scene, providing positioning guarantee for the execution of subsequent logistics unmanned vehicle perception and planning tasks. The logistics unmanned vehicle is also equipped with a binocular camera. By deploying a lightweight target detection algorithm based on the improved YOLOv5, it can recognize specific targets in real time during high-speed driving, complete the recognition of a series of traffic signals such as traffic lights and start-stop signs during driving, and take corresponding actions. At the same time, in order to ensure fast and safe driving in express delivery tasks, the global navigation based on the Hybrid A* algorithm and the local path planning algorithm based on the TD3 reinforcement learning framework are combined with the laser point cloud map, and the local perception information perceived by the laser radar is used. While being able to avoid pedestrian obstacles in time, the real-time response of global navigation and local path planning is realized, efficient optimization and intelligent decision-making are achieved, and the navigation system of the logistics unmanned vehicle has a higher efficiency and intelligence level.
[0135] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0136] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.
Claims
1. A method for express delivery by unmanned logistics vehicles in a park scenario. It is characterized in that Here are the steps: Step 1: Use the 3D laser radar sensor to obtain 3D laser point cloud information, build a laser point cloud map M of the park scene, and obtain the real-time position information P of the logistics unmanned vehicle. c ; Step 2: Based on the constructed laser point cloud map M and the global target point G, a global path planning algorithm hybridA* is used to generate a global trajectory P g ; Step 3: Based on the constructed laser point cloud map M, combined with the real-time pose information P c , calling the YOLOv5 model to detect traffic signs T on the image I of the park scene during driving; Step 4: Based on the real-time detected traffic sign T, combined with traffic rules and cargo pickup and delivery instructions, publish the local target point W l , change the instant driving strategy S, adopt the local path planning based on TD3 reinforcement learning, and generate a new local path trajectory P l ; Step 5: Execute the local path P l Afterwards, execute the global path P g , and continue to collect perception information until reaching the end point G.
2. According to the method for express delivery of unmanned logistics vehicles in a park scenario according to claim 1, It is characterized in that Step 1: Use the 3D laser radar sensor to obtain 3D laser point cloud information, build a laser point cloud map M of the park scene, and obtain the real-time position information P of the unmanned vehicle. c , the specific method is: The 3D laser point cloud information is obtained by using the 3D laser radar sensor, and the LeGO-LOAM algorithm is used to build the laser point cloud map M of the park scene. The laser point cloud map P contains static obstacles O. s and traffic signs T; the ndt matching positioning algorithm is used to match the previously established laser point cloud map M to obtain the current location information P of the logistics unmanned vehicle c .
3. According to the method for express delivery of unmanned logistics vehicles in a park scenario according to claim 1, It is characterized in that Step 3: Based on the constructed laser point cloud map M, combined with the real-time pose information P c , calling the YOLOv5 model, detecting the traffic sign T on the image I of the park scene during driving, extracting the traffic sign features h under the park driving rules, and providing perception information for the express delivery task of the logistics unmanned vehicle. The specific method is as follows: The YOLOv5 model consists of three parts: Backbone, Neck, and detection head. Backbone uses the CSPDarkNet53 network to extract features from the real-time image I collected by the monocular camera during the driving of the logistics unmanned vehicle to obtain multi-scale feature information n 1 ; Neck uses PAFPN to fuse multi-scale features of real-time images extracted from Backbone 1 ; The detection head outputs multi-scale features n according to Neck 1 Predict the target category c and set the prediction box b. Backbone replaces the C3 module with the CSPCM module to achieve the purpose of optimizing the Backbone network structure. At the same time, the CA attention mechanism for specific targets is introduced to focus attention on the target area R. After detecting the traffic sign T, the detection area is solidified and the weight of the key feature points is increased to enhance the model's recognition accuracy of the traffic sign T. (1) Backbone replaces the C3 module with the CSPCM module to achieve the purpose of optimizing the Backbone network structure; The input image I of the logistics unmanned vehicle is first sent to CSPDarkNet53 for preliminary multi-scale feature extraction, and then these multi-scale features n 1 The feature information is transferred to the CSPCM module, which enhances the feature expression capability by stacking ConvMixer layers and 1x1 convolution layers, and performs feature fusion and splicing in the channel dimension. 2 It is then input into the Neck part for multi-scale feature fusion and then sent to the detection head for accurate prediction of the target category c and generation of the prediction box b. (2) Introducing a CA attention mechanism for specific targets to focus attention on the target area R. After detecting the traffic sign T, the detection area is solidified and the weights of key feature points are increased, thereby enhancing the recognition accuracy of the model for the traffic sign T. The feature map output by CSPDarkNet53 is used as the input of the CA attention module. Then, the global statistical information in the feature map is extracted through global average pooling, and the fully connected layer is used to transform and reduce the dimension of the pooled global statistical information. Subsequently, nonlinearity is introduced through the activation function to further activate and refine the feature representation. Finally, the key features of the target area R are highlighted through feature enhancement, while the feature response of the non-target area is suppressed to generate a weighted feature map and send it to the Neck part. Among them, the key features of the target area R are highlighted by feature enhancement, while the feature response of the non-target area is suppressed to generate a weighted feature map. Specifically, according to the relative distance information between the logistics unmanned vehicle and the target, the weight of the feature point is dynamically adjusted, and the close feature points are given higher weights, while the weights of the long-distance feature points are gradually reduced. The hyperbolic tangent function used by the weight value function is as follows: p = tanh(d) Among them, d is the relative distance between the logistics unmanned vehicle and the target, and p is the weight value of each feature point.
4. According to claim 3, the unmanned vehicle express delivery method for logistics in the park scene, It is characterized in that Before using the YOLOv5 model to detect targets in input images or video frames, the target detection frame correction strategy is used to perform geometric correction on the detection frame. The corrected target detection frame makes the static feature points appear on the corresponding optical flow epipolar line l1, while the dynamic target has a large deviation from the epipolar line l1. The specific method is as follows: The target detection frame correction strategy comprehensively collects the posture information, position and speed information, and the distance information between the logistics unmanned vehicle and the target point, constructs the posture state conversion matrix, and corrects the target bounding box. The target detection frame geometry correction function is as follows: x'=f(Δx)+x y'=f(Δy)+y in'=p w ·(Δw)·w h’=p h ·(Δh)·h Among them, the coordinates of the center point of the original detection frame are (x, y), the width and height are w and h respectively, and the offsets output by the YOLOv5 model are Δx, Δy, Δw, Δh. Through the target detection frame geometry correction function, the coordinate values of the corrected target detection frame are: x’ 1 =T{11}·x 1 +T{12}·y 1 +T{13} y’ 1 =T{21}·x 1 +T{22}·y 1 +T{23} x’ 2 =T{11}·x 2 +T{12}·y 2 +T{13} y’ 2 =T{21}·x 2 +T{22}·y 2 +T{23} x’ 3 =T{11}·x 3 +T{12}·y 3 +T{13} y’ 3 =T{21}·x 3 +T{22}·y 3 +T{23} x’ 4 =T{11}·x 4 +T{12}·y 4 +T{13} y’ 4 =T{21}·x 4 +T{22}·y 4 +T{23} Among them, (x1, y1), (x2, y2), (x3, y3), (x4, y4) are the coordinate values of the four vertices of the detection box, T{n 1 n 2 } is point n 1 To point n 2 On this basis, the YOLOv5 model is used to detect the target area and distinguish the background area and traffic sign T.
5. The method for express delivery of unmanned logistics vehicles in a park scenario according to claim 3 or 4, It is characterized in that Use the YOLOv5 model to detect targets on the input image or video frame, obtain a series of initially identified traffic sign T bounding boxes and corresponding confidence scores, and then execute the pseudo target point removal algorithm. The specific method is as follows: Use the YOLOv5 model to detect the target of the input image or video frame, obtain a series of initially recognized traffic sign T bounding boxes and corresponding confidence scores, and then execute the pseudo target point elimination algorithm. Use the optical flow estimation technology to track the motion trajectory of the target area feature points, that is, the traffic sign T, analyze the kinematic model of the target area feature points in continuous frames, establish the dynamic motion model of each feature point, compare it with the preset threshold, identify the abnormal points that do not meet the kinematic characteristics of the traffic sign T, and mark them as potential pseudo target points; Calculate the relative average speed of the traffic sign T to the vehicle body and construct the kinematic model as follows: Among them, E x 、E y The components of the relative average velocity of the target area relative to the vehicle body along the x and y directions, N is the number of feature points for calculating the pose information in the target area, and u j 、v j Calculate the components of the feature point motion velocity along the x and y directions for the vehicle body relative to the posture information.
6. The method for express delivery of unmanned logistics vehicles in a park scenario according to claim 1, It is characterized in that Step 4: Based on the real-time detected traffic sign T, combined with traffic rules and cargo pickup and delivery instructions, publish the local target point W l , change the instant driving strategy S, adopt the local path planning based on reinforcement learning, and generate a new local path trajectory P l , the specific method is: Logistics unmanned vehicle environment information obtains local environmental information through laser radar e , and the polar coordinate information of the target relative to the logistics unmanned vehicle in polar coordinates Combination, where d represents the Euclidean distance between the current position of the unmanned vehicle and the target position, Represents the interpolation angle between the current angle of the unmanned vehicle and the target position, s e and r The combined data will be used as input state s in the action network of TD3; The state s is input into the action network of the reinforcement learning network. The action network consists of two fully connected layers (FC). After each fully connected layer, it is activated by ReLU (activation function). The activation layer is connected to the output layer, and the two action parameters are output, which are the linear speed a of the logistics unmanned vehicle and 1 and angular velocity a 2 ; Then, the output value of the output layer is controlled to (-1,1) through tanh (hyperbolic tangent function), and then action a is applied to the logistics unmanned vehicle. The specific formula is: Among them, v max is the maximum linear velocity, ω max is the maximum angular velocity; After the action network obtains the action of the unmanned vehicle, the Q value of the action network is evaluated by two evaluation networks. The two evaluation networks have the same structure. The state s and action a are used as the input of the evaluation network. The state s is input to the fully connected layer of the evaluation network. The fully connected layer is activated by ReLU and outputs L s Output L s and actions will be input into independent transformation fully connected layers τ with the same size 1 and τ 2 The combination is as follows: Among them, L C is the output of the combined fully connected layer (CFC), and They are the fully connected layers τ 1 and τ 2 The weight of is the fully connected layer τ 2 The bias of the network is rectified; ReLU is applied to the combined layer and 1 parameter is output to represent the Q value; the smallest Q value of the two evaluation networks is selected as the final evaluation parameter output to limit the overestimation of the Q value by the actor-critic network; The reward function is as follows: The state-action pair (s t ,a t ) depends on three conditions. If at the current time step D t The distance to the target is less than the threshold η D , then a positive reward r is used g ; If a collision is detected, a negative reward r is used c ; If none of them are satisfied, an immediate reward is used based on the current linear velocity v and angular velocity ω; In order to guide the navigation strategy towards the set goal and obtain different parameter values, the parameters of the two are updated delayedly, and the delayed attribute reward method is used to obtain a more stable reward value. The specific formula is as follows: Among them, n is the number of previous steps to update the reward. The positive reward is not only determined by the state-action when the goal is reached, but also decreases in the last n steps before reaching the goal. Through the constraint of the reward function on the network parameters, the reinforcement learning network learns that the logistics unmanned vehicle moves from the current position to the local target point W. l The local path trajectory is merged and updated into the final global trajectory until the global target point G is reached.
7. A logistics unmanned vehicle express delivery system in a park scenario, It is characterized in that Implement the method for express delivery of logistics unmanned vehicles in a park scenario as described in any one of claims 1-6 to achieve express delivery of logistics unmanned vehicles in a park scenario.
8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for express delivery of logistics unmanned vehicles in a campus scenario as described in any one of claims 1 to 6 is implemented to achieve express delivery of logistics unmanned vehicles in a campus scenario.
9. A computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for express delivery of logistics unmanned vehicles in a campus scenario as described in any one of claims 1 to 6 is implemented to achieve express delivery of logistics unmanned vehicles in a campus scenario.
Citation Information
Cited By
Automatic driving vehicle pick-up point generation system based on space mapping and environment remodeling
CN119625231A
Continuous space path planning method based on deep reinforcement learning in unknown environment
CN120085555A
