Target pedestrian trajectory prediction method in combination with motion intention
By combining a low-cost monocular camera with YOLO, BlazePose, Transformer, and GCN networks, the problems of reliance on high-cost 3D sensors and insufficient prediction accuracy in complex environments are solved, achieving high-precision prediction of target pedestrian trajectories and improving the robustness and accuracy of robot autonomous following.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ROBOTICS RESEARCH CENTER OF YUYAO CITY
- Filing Date
- 2025-12-22
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies rely on high-cost 3D sensors and have insufficient prediction accuracy in complex environments. They lack target trajectory prediction methods that utilize prior information about pedestrian movement, and publicly available datasets lack annotations of real 3D locations and skeleton data.
By employing a low-cost monocular camera combined with YOLO, BlazePose, Transformer, and GCN networks, and through target detection, skeleton joint key point extraction, historical trajectory sequence analysis, and an improved ByteTrack algorithm, the trajectory prediction of the target pedestrian is achieved.
Achieving high-precision trajectory prediction on low-cost hardware improves robustness and accuracy in complex scenarios, reduces reliance on high-cost sensors, and enhances the model's generalization ability in diverse scenarios.
Smart Images
Figure CN121963133A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of robot vision and robotics technology, specifically relating to a method for predicting the trajectory of a target pedestrian by incorporating motion intent. Background Technology
[0002] With the rapid development of robotics technology, autonomous following has become a key technology in service robots, logistics robots, and other fields. Achieving accurate prediction of target pedestrian trajectories is a core aspect of improving robot following performance. Current technologies largely rely on 3D sensors (such as depth cameras and LiDAR) to acquire high-precision 3D data. While these technologies offer high accuracy under ideal conditions, they still present the following technical challenges:
[0003] (1) The hardware costs are high and there are performance bottlenecks in complex environments.
[0004] (2) The target trajectory prediction method only uses historical location sequence data and does not make reasonable use of the prior information of pedestrian movement, which leads to a decrease in prediction accuracy when the pedestrian movement direction changes abruptly.
[0005] (3) Most existing public datasets lack annotation information that simultaneously contains real 3D position and skeleton data, which limits the development of trajectory prediction methods that fuse pose information.
[0006] Therefore, there is an urgent need for a new method to achieve high-precision target pedestrian trajectory prediction from single-frame images using a low-cost monocular camera. Summary of the Invention
[0007] To address the aforementioned technical problems in existing technologies, the purpose of this invention is to achieve high-precision target pedestrian trajectory prediction from a single frame image using a low-cost monocular camera, providing a new approach for robot autonomous following and interactive perception, reducing hardware dependence and improving robustness in complex scenarios. The technical solution is as follows:
[0008] A method for predicting the trajectory of a target pedestrian based on incorporating motion intent includes the following steps:
[0009] Step 1: Acquire the image, use YOLO to detect the target human body, and output the target detection bounding box;
[0010] Step 2: Zoom in on the target detection box to obtain the region of interest (ROI) in the image, and extract the key points of the two-dimensional skeleton joints in the ROI using the BlazePose network;
[0011] Step 3: Calculate the target's 3D position relative to the camera using the monocular camera projection model and based on the key points of the 2D skeleton joints, and store it in the historical trajectory sequence;
[0012] Step 4: By fusing the prediction network of Transformer and GCN, obtain the spatiotemporal information of the target's historical trajectory sequence and the key points of the two-dimensional skeleton joints, and predict the two-dimensional Gaussian distribution parameters of the target's future trajectory;
[0013] Step 5: Use the improved ByteTrack algorithm for target tracking. Update the target detection box based on the target detection box described in Step 1 and the output of the predicted future trajectory of the target. Provide the updated detection box to Step 2 for the next frame processing.
[0014] Step 6: Integrate the 2D target detection box, 3D target position, and predicted motion trajectory to output structured results.
[0015] Furthermore, in step 1, the image is a single-frame color image with a resolution of 640×480 captured by a monocular camera.
[0016] Furthermore, in step 1, the YOLO detection of the target human body involves using the YOLO target detection network to perform feedforward processing on the input image and outputting a detection box for the target pedestrian, the expression of which is:
[0017]
[0018] in, Indicates the center coordinates of the detection box. Indicates the width of the detection box. This indicates the height of the detection frame.
[0019] Furthermore, in step 2, enlarging the target detection box involves scaling up the target detection box obtained in step 1 by a scaling factor to obtain the ROI region. The expression of the enlarged detection box is:
[0020]
[0021] in, Indicates the scaling factor;
[0022] The extraction of key points of the two-dimensional skeleton joints in the ROI region is performed using the BlazePose network. The key points of a two-dimensional skeleton joint are expressed as follows:
[0023]
[0024] in, Indicates the first The two-dimensional coordinates of each joint in the image coordinate system.
[0025] Furthermore, the scaling factor value .
[0026] Furthermore, in step 3, the specific process of calculating the target's three-dimensional position relative to the camera and storing it in the historical trajectory sequence is as follows:
[0027] (1) The preset camera intrinsic parameter matrix is as follows: Its expression is:
[0028]
[0029] in, Indicates focal length. Indicates the coordinates of the principal point;
[0030] 3D joints The relationship between projection onto the image plane is as follows:
[0031]
[0032] in, Indicates the scale factor;
[0033] (2) Set the standard length of the main human skeleton. Based on the prior length constraints of the main human skeleton, solve the three-dimensional position of the target pedestrian relative to the camera by minimizing the joint objective function of the two-dimensional projection error and the skeleton length constraint. The calculation expression is as follows:
[0034]
[0035] in, This represents the prior length constraint of the human skeleton. Represents the projection function. This represents the estimated bone length. Represents the regularization coefficient;
[0036] (3) Set the current time The position is stored in the historical trajectory sequence, expressed as:
[0037]
[0038] in, Indicates the size of the history window.
[0039] Furthermore, the standard lengths of the main human skeletal structures include: shoulder width, upper arm length, and forearm length.
[0040] Furthermore, in step 4, the specific process for determining the two-dimensional Gaussian distribution parameters of the predicted target's future trajectory is as follows:
[0041] (1) Input the historical trajectory sequence and the key points of the two-dimensional skeleton joints into the prediction network that fuses Transformer and GCN;
[0042] (2) The GCN prediction network encodes the key points of the two-dimensional skeleton joints, and models the human skeleton as a graph structure, the expression of which is:
[0043]
[0044] in, Represents the set of key points. Represents the set of edges connecting the skeleton;
[0045] (3) GCN extracts the spatial topological features of human pose through graph convolution and generates a pose feature representation. Its calculation expression is as follows:
[0046]
[0047] in, This represents the adjacency matrix with added self-loops. Degree matrix, Indicates learnable parameters, Indicates the activation function;
[0048] (4) The Transformer model uses the historical trajectory sequence as the key and value tokens, and the skeleton key points encoded by the GCN prediction network as the query token. It obtains the spatiotemporal dependencies in the time series through the attention mechanism. Its calculation expression is:
[0049]
[0050] in, This represents the query token. Indicates key, Indicates the value token;
[0051] (5) The two-dimensional Gaussian distribution of the predicted future trajectory of the target is output through the fully connected layer, and its expression is:
[0052]
[0053] in, This represents the mean vector of the predicted trajectory. Let represent the covariance matrix.
[0054] Furthermore, in step 5, the specific process of updating the target detection box is as follows:
[0055] (1) Define the state vector of the Kalman filter as follows: Its expression is:
[0056]
[0057] in, Indicates the center position of the detection frame. Indicates the aspect ratio. Indicates altitude, These represent their corresponding speeds;
[0058] (2) The target detection box described in step 1 and the predicted future trajectory of the target are fused together, and the expression is:
[0059]
[0060] in, Represents the state transition matrix. This represents a function that converts the predicted trajectory into state increments. Indicates the fusion weights;
[0061] (3) The Hungarian algorithm is used to perform association matching between the detection box and the trajectory, and the updated target detection box is output. .
[0062] Furthermore, in step 6, the structured output results include: the target detection box of the current frame, the three-dimensional position of the target in the camera coordinate system, and the predicted trajectory distribution of future frames.
[0063] Beneficial effects: (1) The entire process design is realized under low-cost hardware conditions. The entire process from target detection to trajectory prediction can be completed by relying only on a monocular camera, which reduces the dependence on high-cost 3D sensors and provides an economical and efficient solution for robot autonomous target following. (2) By extracting the key points of human skeleton joints and encoding the spatial topological features of human posture by combining GCN, implicit modeling of pedestrian movement intention is realized, which improves the accuracy and robustness of pedestrian trajectory prediction. (3) By integrating the predicted trajectory into the target tracking module, the accuracy and robustness of target tracking in crowded and complex dynamic scenes are improved. (4) By adopting a network architecture that integrates Transformer and GCN, Transformer captures the temporal dependency of historical trajectories and GCN extracts the spatial topological features of human posture, which realizes the effective fusion of spatiotemporal information. Through simulation data augmentation strategy, training data is generated in diverse terrain scenes, (5) which improves the generalization ability of the model in complex real scenes. Attached Figure Description
[0064] Figure 1 This is a flowchart of a method for predicting the trajectory of a target pedestrian that incorporates motion intent, according to the present invention.
[0065] Figure 2 This is an example diagram of the three-dimensional human skeleton joint projection of the present invention;
[0066] Figure 3 This is a diagram of the prediction network structure that integrates Transformer and GCN in this invention.
[0067] Figure 4 This is a schematic diagram illustrating target trajectory prediction based on motion intent in a warehousing scenario according to the present invention. Detailed Implementation
[0068] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0069] This invention discloses a method for predicting the trajectory of a target pedestrian by incorporating motion intent. Using autonomous pedestrian following by an intelligent robot as an example, the method utilizes data collected by a monocular camera and combines target detection, skeleton extraction, position estimation, trajectory prediction, and target tracking modules to verify its effectiveness. The preliminary preparation steps are as follows:
[0070] (1) Create a dataset
[0071] We selected the Human3.6M and CMU Motion Capture datasets, which contain human joint annotations. However, these datasets are single-scene datasets that involve motion in a confined space, and they lack scene diversity.
[0072] To collect pedestrian motion data in complex scenes, data augmentation was performed on the above dataset. The Isaac Sim high-fidelity simulator was selected, and motion stylization techniques were used to move the Humanoid model in diverse terrain scenes, resulting in a dataset with more complex scenes and more diverse motion.
[0073] The newly generated dataset contains the spatial 3D coordinates and joint positions of the target pedestrian, as well as the 3D position of the camera. Based on the Rodrigues rotation formula, the corresponding human skeletal joint positions are calculated forward from the 3D joint positions of the Humanoid model:
[0074]
[0075] in, Indicates the axis of rotation. Indicates the rotation angle. This represents an antisymmetric matrix.
[0076] Based on the 3D positions of the camera and the pedestrian, calculate the position of the target and its 3D joints in the camera coordinate system:
[0077]
[0078] in and Let represent the rotation matrix and translation vector from the world coordinate system to the camera coordinate system, respectively.
[0079] Based on the pinhole camera imaging model, the ground truth values of the target detection box and skeleton joint positions can be obtained, for 3D points. Its two-dimensional projected coordinates are:
[0080]
[0081] A dataset including 3D joint annotations, 2D joint key points, and target motion trajectory was established.
[0082] (2) Training the model
[0083] Applying noise to the target detection bounding box simulates instability during the detection process; its expression is:
[0084]
[0085] in, This represents the standard deviation of noise.
[0086] To randomize target pedestrians of different body types by applying a certain scale change to the skeleton, the expression is as follows:
[0087]
[0088] in, This represents the random scaling factor.
[0089] To simulate lens calibration error, the camera intrinsic parameters are randomized to a certain extent; the expression is as follows:
[0090]
[0091] The experiments of this invention were conducted on a cloud platform server. The preprocessed training images were resized to 320×240 pixels and bilinear interpolation was used. The loss function, calculated by combining object detection bounding box prediction and object motion trajectory prediction, is as follows:
[0092]
[0093] in, Indicates the loss of the detection frame. This represents the trajectory prediction loss. This represents the attitude estimation loss. This represents the balance coefficient.
[0094] like Figure 1 The specific steps of the target pedestrian trajectory prediction method combining movement intention of the present invention are as follows:
[0095] S1. Acquire a single-frame color image using a monocular camera, use YOLO to detect the target human body, and output the target detection bounding box;
[0096] A single-frame color image with a resolution of 640×480 was captured using a monocular camera. In the actual implementation, the image size was adjusted to 320×240 to accommodate subsequent processing requirements. The input image was in RGB format to ensure it contained complete information about the human target, providing basic data for subsequent detection.
[0097] The system performs feedforward processing on the input image using the YOLO object detection network. YOLO divides the input image into an S×S grid, and each grid predicts B bounding boxes and their confidence scores. For the pedestrian detection task, boxes with confidence scores higher than a threshold are selected. The detection results are used to output the bounding box of the target pedestrian.
[0098] YOLO target human detection works by using the YOLO object detection network to perform feedforward processing on the input image and output the detection bounding box of the target pedestrian. Its expression is:
[0099]
[0100] in, Indicates the center coordinates of the detection box. Indicates the width of the detection box. This indicates the height of the detection frame.
[0101] S2. Appropriately enlarge the BoundingBox to obtain the Region of Interest (ROI) of the image, and extract the key points of the two-dimensional skeleton joints in the ROI using BlazePose.
[0102] BoundingBox by scaling factor Enlarge the area appropriately to obtain the Region of Interest (ROI). The expression of the enlarged detection box is:
[0103]
[0104] in, Represents the scaling factor, and ;
[0105] The extraction of key points of the two-dimensional skeleton joints in the ROI region is performed using the BlazePose network. The key points of a two-dimensional skeleton joint are expressed as follows:
[0106]
[0107] in, Indicates the first The two-dimensional coordinates of each joint in the image coordinate system.
[0108] BlazePose is a lightweight human pose estimation network capable of detecting 33 human keypoints in real time, including joint positions on the body, face, and hands. The output keypoint set... This provides input for subsequent 3D position estimation.
[0109] S3, such as Figure 2 As shown, a monocular camera projection model is used to estimate the target's three-dimensional position relative to the camera based on the two-dimensional skeleton joint key points, and the result is stored in the historical trajectory sequence.
[0110] Based on a single-lens camera pinhole imaging model, this invention estimates the 3D position of a target relative to the camera from key points of a 2D skeleton joint. Since a single-lens camera cannot directly acquire depth information, this invention utilizes prior length constraints of the human skeleton and solves for the target's 3D position through an optimization algorithm. Specifically, standard lengths of the main human skeletons are set (e.g., shoulder width, upper arm length, forearm length, etc.), and the 3D position of the pedestrian in the camera coordinate system is solved by minimizing the joint objective function of 2D projection error and skeleton length constraints. The specific process is as follows:
[0111] (1) The preset camera intrinsic parameter matrix is as follows: Its expression is:
[0112]
[0113] in, Indicates focal length. Indicates the coordinates of the principal point;
[0114] 3D joints The relationship between projection onto the image plane is as follows:
[0115]
[0116] in, Indicates the scale factor;
[0117] (2) Set the standard length of the main human skeleton. Based on the prior length constraints of the main human skeleton, solve the three-dimensional position of the target pedestrian relative to the camera by minimizing the joint objective function of the two-dimensional projection error and the skeleton length constraint. The calculation expression is as follows:
[0118]
[0119] in, This represents the prior length constraint of the human skeleton. Represents the projection function. This represents the estimated bone length. Represents the regularization coefficient;
[0120] (3) Set the current time The position is stored in the historical trajectory sequence, expressed as:
[0121]
[0122] in, Indicates the size of the history window.
[0123] S4, such as Figure 3 As shown, a prediction network that integrates Transformer and GCN is used to capture the spatiotemporal information of the target's historical trajectory sequence and two-dimensional skeleton joint key points, and to predict the two-dimensional Gaussian distribution of the target's future trajectory.
[0124] The specific process of predicting the two-dimensional Gaussian distribution parameters of the target's future trajectory is as follows:
[0125] (1) Input the historical trajectory sequence and the key points of the two-dimensional skeleton joints into the prediction network that fuses Transformer and GCN;
[0126] (2) The GCN prediction network encodes the key points of the two-dimensional skeleton joints, and models the human skeleton as a graph structure, the expression of which is:
[0127]
[0128] in, Represents the set of key points. Represents the set of edges connecting the skeleton;
[0129] (3) GCN extracts the spatial topological features of human pose through graph convolution and generates a pose feature representation. Its calculation expression is as follows:
[0130]
[0131] in, This represents the adjacency matrix with added self-loops. Degree matrix, Indicates learnable parameters, Indicates the activation function;
[0132] (4) The Transformer model uses the historical trajectory sequence as the key and value tokens, and the skeleton key points encoded by the GCN prediction network as the query token. It obtains the spatiotemporal dependencies in the time series through the attention mechanism. Its calculation expression is:
[0133]
[0134] in, This represents the query token. Indicates key, Indicates the value token;
[0135] (5) The two-dimensional Gaussian distribution of the predicted future trajectory of the target is output through the fully connected layer, and its expression is:
[0136]
[0137] in, This represents the mean vector of the predicted trajectory. Let represent the covariance matrix.
[0138] S5. Use the improved ByteTrack for target tracking. ByteTrack is a high-efficiency multi-target tracking algorithm that improves tracking robustness by associating high-confidence and low-confidence detection boxes. Update the target detection boxes using the outputs of the target detection boxes described in S1 and the predicted future trajectory of the target. The specific process is as follows:
[0139] (1) Define the state vector of the Kalman filter as follows: Its expression is:
[0140]
[0141] in, Indicates the center position of the detection frame. Indicates the aspect ratio. Indicates altitude, These represent their corresponding speeds;
[0142] (2) The target detection box described in S1 and the predicted future trajectory of the target are fused together, and the expression is as follows:
[0143]
[0144] in, Represents the state transition matrix. This represents a function that converts the predicted trajectory into state increments. Indicates the fusion weights;
[0145] (3) The Hungarian algorithm is used to perform association matching between the detection box and the trajectory, and the updated target detection box is output. .
[0146] The updated detection box is provided to step S2 for the next frame processing.
[0147] S6. The system integrates 2D detection boxes, 3D target positions, and predicted motion trajectories, outputting structured results for autonomous target following by the robot. Output data includes the target detection box in the current frame, the target's 3D position in the camera coordinate system, and the predicted trajectory distribution for several future frames.
[0148] like Figure 4 As shown, the scenario of this invention is set as an AMR (Autonomous Mobile Robot) in a warehousing and logistics center or an autonomous driving assistance system at a pedestrian crossing in the city. In such scenarios, the motion state of the target pedestrian undergoes a non-linear abrupt change, that is, the target pedestrian often suddenly changes the direction of travel without warning (for example, the pedestrian was originally walking in a straight line, suddenly found a lost item and stopped abruptly and turned to run, or suddenly cut into the robot's path to avoid an obstacle).
[0149] The intention from pose recognition and Transformer attention mechanism correction in this invention are as follows: When a pedestrian is about to suddenly change direction and run, although their center of mass (position) has not yet shifted significantly, their limb joints (such as head turning, shoulder tilting, and leg stride angles) have already undergone drastic changes in the first frame image. In S4 of this invention, the GCN (Graph Convolutional Network) directly performs topological encoding on the key points of the two-dimensional skeleton joints. Even if the historical trajectory is displayed as a straight line, the skeleton features (Feature Token) extracted by the GCN in the current frame will contain strong semantic information of "turning" or "running". In the fusion layer of the Transformer, the skeleton pose features (Query) of the current frame are given a higher weight than the historical trajectory. This means that the system can instantly overturn the old trajectory prediction and directly output a new predicted trajectory that conforms to human biomechanics within 1-2 frames (about 30-60ms) of capturing the abrupt change in pose.
[0150] This invention is mainly applied to autonomous robots, assisted driving, and intelligent security scenarios.
Claims
1. A method for predicting the trajectory of a target pedestrian based on combined motion intention, characterized in that, Includes the following steps: Step 1: Acquire the image, use YOLO to detect the target human body, and output the target detection bounding box; Step 2: Zoom in on the target detection box to obtain the region of interest (ROI) in the image, and extract the key points of the two-dimensional skeleton joints in the ROI using the BlazePose network; Step 3: Calculate the target's 3D position relative to the camera using the monocular camera projection model and based on the key points of the 2D skeleton joints, and store it in the historical trajectory sequence; Step 4: By fusing the prediction network of Transformer and GCN, obtain the spatiotemporal information of the target's historical trajectory sequence and the key points of the two-dimensional skeleton joints, and predict the two-dimensional Gaussian distribution parameters of the target's future trajectory; Step 5: Use the improved ByteTrack algorithm for target tracking. Update the target detection box based on the target detection box described in Step 1 and the output of the predicted future trajectory of the target. Provide the updated detection box to Step 2 for the next frame processing. Step 6: Integrate the 2D target detection box, 3D target position, and predicted motion trajectory to output structured results.
2. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 1, the image is a single-frame color image with a resolution of 640×480 captured by a monocular camera.
3. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 1, the YOLO detection of the target human body involves using the YOLO target detection network to perform feedforward processing on the input image and outputting a detection box for the target pedestrian. The expression for this box is: in, Indicates the center coordinates of the detection box. Indicates the width of the detection box. This indicates the height of the detection frame.
4. The method for predicting the trajectory of a target pedestrian based on combined motion intent as described in claim 1, characterized in that, In step 2, enlarging the target detection box involves scaling up the target detection box obtained in step 1 by a scaling factor to obtain the Region of Interest (ROI). The expression for the enlarged detection box is: in, Indicates the scaling factor; The extraction of key points of the two-dimensional skeleton joints in the ROI region is performed using the BlazePose network. The key points of a two-dimensional skeleton joint are expressed as follows: in, Indicates the first The two-dimensional coordinates of each joint in the image coordinate system.
5. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 4, characterized in that, The scaling factor value .
6. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 3, the specific process of calculating the three-dimensional position of the target relative to the camera and storing it in the historical trajectory sequence is as follows: (1) The preset camera intrinsic parameter matrix is as follows: Its expression is: in, Indicates focal length. Indicates the coordinates of the principal point; 3D joints The relationship between projection onto the image plane is as follows: in, Indicates the scale factor; (2) Set the standard length of the main human skeleton. Based on the prior length constraints of the main human skeleton, solve the three-dimensional position of the target pedestrian relative to the camera by minimizing the joint objective function of the two-dimensional projection error and the skeleton length constraint. The calculation expression is as follows: in, This represents the prior length constraint of the human skeleton. Represents the projection function. This represents the estimated bone length. Represents the regularization coefficient; (3) Set the current time The position is stored in the historical trajectory sequence, expressed as: in, Indicates the size of the history window.
7. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 6, characterized in that, The standard lengths of the main human bones include: shoulder width, upper arm length, and forearm length.
8. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 4, the specific process of obtaining the two-dimensional Gaussian distribution parameters of the predicted future trajectory of the target is as follows: (1) Input the historical trajectory sequence and the key points of the two-dimensional skeleton joints into the prediction network that fuses Transformer and GCN; (2) The GCN prediction network encodes the key points of the two-dimensional skeleton joints, and models the human skeleton as a graph structure, the expression of which is: in, Represents the set of key points. Represents the set of edges connecting the skeleton; (3) GCN extracts the spatial topological features of human pose through graph convolution and generates a pose feature representation. Its calculation expression is as follows: in, This represents the adjacency matrix with added self-loops. Degree matrix, Indicates learnable parameters, Indicates the activation function; (4) The Transformer model uses the historical trajectory sequence as the key and value tokens, and the skeleton key points encoded by the GCN prediction network as the query token. It obtains the spatiotemporal dependencies in the time series through the attention mechanism. Its calculation expression is: in, This represents the query token. Indicates key, Indicates the value token; (5) The two-dimensional Gaussian distribution of the predicted future trajectory of the target is output through the fully connected layer, and its expression is: in, This represents the mean vector of the predicted trajectory. Let represent the covariance matrix.
9. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 5, the specific process of updating the target detection box is as follows: (1) Define the state vector of the Kalman filter as follows: Its expression is: in, Indicates the center position of the detection frame. Indicates the aspect ratio. Indicates altitude, These represent their corresponding speeds; (2) The target detection box described in step 1 and the predicted future trajectory of the target are fused together, and the expression is: in, Represents the state transition matrix. This represents a function that converts the predicted trajectory into state increments. Indicates the fusion weights; (3) The Hungarian algorithm is used to perform association matching between the detection box and the trajectory, and the updated target detection box is output. .
10. The method for predicting the trajectory of a target pedestrian based on combined motion intention as described in claim 1, characterized in that, In step 6, the structured output results include: the target detection box of the current frame, the three-dimensional position of the target in the camera coordinate system, and the predicted trajectory distribution of future frames.