A robot active safety obstacle avoidance method based on human behavior understanding and space-time trajectory prediction

Abstract

The application relates to the technical field of robot active safety obstacle avoidance, and discloses a robot active safety obstacle avoidance method based on human behavior understanding and space-time trajectory prediction, which comprises the following steps: collecting a test RGB video stream and constructing a multi-task data set; end-to-end training of a multi-task parallel network; when online obstacle avoidance is performed, the multi-task parallel network extracts deep space geometric feature vectors, which are distributed in parallel to three task branches of limb segmentation, behavior classification and trajectory prediction; the current behavior category of a pedestrian is determined, and a category probability distribution is injected into the trajectory prediction branch to calculate a future 5-second full-limb motion trajectory; the trajectory is converted into a three-dimensional physical space trajectory coordinate sequence; the sequence is updated in real time into a space-time dynamic cost map, a dynamic safety protection boundary is updated, an optimal detour or deceleration path is calculated, and active obstacle avoidance is performed; the application has the advantages of high long-period trajectory prediction accuracy and good real-time response of an embedded end, and effectively reduces the accumulation of previous classification errors and trajectory drift.

Description

Technical Field

[0001] This invention relates to the field of active safety obstacle avoidance technology for robots, specifically to an active safety obstacle avoidance method for robots based on human behavior understanding and spatiotemporal trajectory prediction. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, mobile robots (such as commercial service robots, warehousing and logistics AGVs, and hospital guidance robots) are integrating into people's daily lives and work environments at an unprecedented pace. In dynamic and complex environments where humans and machines coexist, how mobile robots can achieve safe, efficient, and socially compliant obstacle avoidance and navigation is a key technology for ensuring human-machine safety and improving the quality of robot services.

[0003] Currently, traditional obstacle avoidance methods for robots largely rely on perception hardware such as LiDAR or depth cameras, and perform path planning by constructing a spatiotemporal cost map or using artificial potential fields. These methods work well when dealing with static obstacles such as boxes and walls, but they often mechanically treat pedestrians as ordinary "dynamic rigid bodies" or "moving boxes." However, humans exhibit a high degree of randomness, intentionality, and social attributes during movement. Traditional obstacle avoidance methods, lacking a deep understanding of pedestrian behavior and their underlying intentions, often only resort to emergency braking or sharp turns when extremely close to pedestrians. This not only easily leads to frequent "robot freezing problems," significantly reducing navigation efficiency, but also easily causes safety hazards such as human-robot collisions, lacking comfort and social responsibility in human-robot interaction.

[0004] In recent years, although some studies have attempted to introduce deep learning algorithms for pedestrian trajectory prediction to improve the foresight of robot obstacle avoidance, most existing prediction methods rely solely on the pedestrian's historical position coordinates from the past few seconds, ignoring the decisive influence of micro-behavioral characteristics (such as looking down at a phone, suddenly turning around, or stopping to talk) on future trajectories. Furthermore, existing spatiotemporal prediction models generally suffer from high computational complexity and long processing times, making it difficult to meet real-time requirements on embedded platforms for mobile robots with limited computing power. Therefore, how to deeply integrate human behavioral intent understanding with high-precision spatiotemporal trajectory prediction while ensuring computational efficiency, thereby achieving proactive and safe obstacle avoidance for robots, remains a significant challenge in the field of mobile robot navigation. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as treating pedestrians as moving rigid bodies and lacking understanding of limb intentions during robot obstacle avoidance, the amplification of classification errors in existing behavior recognition and trajectory prediction algorithms, and insufficient accuracy in predicting future trajectories when facing complex dynamic behaviors, this invention proposes a robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction. The technical solution is as follows:

[0006] A robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction includes the following steps:

[0007] Step 1: Collect test RGB video stream and capture continuous color images. Jointly annotate the pedestrians in the images and add behavior labels. Record the two-dimensional pixel coordinate sequence of the pedestrian's body center point and key points of the limbs within 5 seconds to construct a multi-task behavior trajectory dataset.

[0008] Step 2: Construct a multi-task parallel network model, including a shared backbone network, a multi-scale feature fusion network, and a multi-task processing branch. Train the model end-to-end using the multi-task behavior trajectory dataset.

[0009] Step 3: Deploy the trained model, capture the original RGB video stream within the moving field of view in real time, extract historical images from the past and present of M consecutive frames, and construct a time-series image group;

[0010] Step 4: Input the time series image group into the shared backbone network, and extract deep spatial geometric feature vectors as basic features through the multi-scale feature fusion network, and input them in parallel into the multi-task processing branches, including the limb segmentation branch, behavior classification branch and trajectory prediction branch;

[0011] Step 5: Combine the calculation results of the multi-task processing branches to generate the predicted trajectory of the pedestrian's two-dimensional image coordinates within the next 5 seconds;

[0012] Step 6: Convert the predicted trajectory of the two-dimensional image coordinates into a three-dimensional physical space trajectory coordinate sequence in the robot world coordinate system, and input it into the navigation planner. The navigation planner dynamically updates the safety protection boundary in the spatiotemporal dynamic cost map, and the local path planner reads and calculates the detour or deceleration path in real time to perform active safety obstacle avoidance.

[0013] Furthermore, the limb segmentation branch calculates and outputs the semantic segmentation mask of the pedestrian's limbs in the current image; the behavior classification branch concatenates the features of M consecutive frames according to the time sequence, analyzes the limb swing pattern through the time self-attention mechanism, and outputs the behavior category determination probability of the current pedestrian; the trajectory prediction branch performs trajectory calculation, introduces the behavior category determination probability, and adaptively corrects the future trajectory movement trend.

[0014] Furthermore, the end-to-end training employs a joint loss function for collaborative optimization, as shown in the following formula:

[0015] ,

[0016] in, The preset weighting coefficients, For human body positioning frame loss, For limb semantic segmentation loss, Loss is categorized by behavior. Loss is calculated for dynamic trajectory prediction.

[0017] Furthermore, the shared backbone network is specifically an improved CSPDarknet backbone network, whose core processing architecture, along the data flow direction, is as follows: a first two-dimensional convolutional layer; a second two-dimensional convolutional layer; a third C3k2 feature extraction module, which contains a series lightweight bottleneck structure and uses the SiLU function as the activation function; a fourth two-dimensional convolutional layer; a fifth C3k2 feature extraction module, which contains a series lightweight bottleneck structure and enables residual shortcut connections; a sixth two-dimensional convolutional layer; a seventh C3k2 feature extraction module, which contains a series lightweight bottleneck structure; an eighth two-dimensional convolutional layer; a ninth C3k2 feature extraction module, which contains a series lightweight bottleneck structure; and a tenth fast spatial pyramid pooling module, which contains a parallel max pooling layer.

[0018] Furthermore, during operation, the shared backbone network distributes the 128-channel feature map output from the seventh layer, the 256-channel feature map output from the ninth layer, and the 256-channel feature map output from the tenth layer in parallel to the multi-scale feature fusion network.

[0019] Furthermore, the multi-scale feature fusion network adopts a path aggregation network architecture with bidirectional fusion from top to bottom and bottom to top. The input end receives the feature map output by the shared backbone network, and deeply integrates deep spatial geometric features with global semantic information through cascaded upsampling and downsampling operations. The feature maps are then distributed in parallel to the multi-task processing branches on demand: high-resolution feature maps rich in spatial topology and edge details are input to the limb segmentation branch; low-resolution feature maps rich in macroscopic dynamic semantics are input to the behavior classification branch; and multi-scale feature maps that take into account both positioning accuracy and motion trends are cascaded and input to the trajectory prediction branch.

[0020] Furthermore, the trajectory prediction branch introduces a feature interaction mechanism for intent constraint weights, mathematically represented as follows: Let the behavior category determination probability vector output by the behavior classification branch be:

[0021] ,

[0022] in, Let be the total number of behavior categories; let the initial latent feature matrix of the trajectory prediction branch be... By constructing an attention mapping matrix The probability of behavior category determination is explicitly injected as a dynamic constraint into the trajectory features, resulting in a fused interaction feature matrix. Represented as:

[0023] ,

[0024] in, It represents the Hadamah accumulation. This represents the Sigmoid activation function.

[0025] Further, step 5 specifically involves the following steps: The behavior classification branch extracts the limb swing amplitude and body center of gravity movement trend features of the pedestrian in the past M consecutive historical images through a temporal cascade encoder; the extracted features are matched with standard action templates in the temporal action primitive state library to calculate the probability of the pedestrian's current behavior category; the decoder of the trajectory prediction branch adaptively calls the kinematic constraint law corresponding to the determined behavior action type to correct and generate the future trajectory.

[0026] Furthermore, the linkage mechanism between behavior categories and safety expansion coefficients in the spatiotemporal dynamic cost map is as follows: if the pedestrian's current behavior is determined to be walking or bending over, then the safety expansion radius around the future trajectory point is set as the baseline value. If the behavior is determined to be running or falling, the safety expansion radius is adjusted to... ,in Increase the weight of collision hazard areas in physical space to force the local path planner to switch to the detour path in advance.

[0027] Beneficial effects

[0028] This invention has the advantages of high accuracy in long-cycle trajectory prediction and good real-time response of embedded terminals. By adaptively constraining and correcting the trajectory divergence direction in the next 5 seconds under different behavioral modes through the interaction head of behavioral intent and temporal trajectory, it effectively reduces the accumulation of pre-classification errors and trajectory drift caused by traditional serial decoupling architecture. The prediction results help the robot respond in advance to routine and sudden behaviors such as standing, walking, running, falling, and bending over, and adjust the safety expansion coefficient in the physical space in conjunction with the prediction, thereby planning an advanced detour path that conforms to social ethics and etiquette, effectively avoiding collision accidents and deadlock phenomena caused by frequent sudden stops of the robot. Attached Figure Description

[0029] Figure 1A flowchart of a robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction;

[0030] Figure 2 This is a diagram of the multi-task parallel network model architecture in an embodiment of the present invention. Detailed Implementation

[0031] 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.

[0032] like Figure 1 As shown, the robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction of the present invention includes:

[0033] Step 1: Test Dataset Collection and Standardized Labeling

[0034] Using a visible light camera mounted on a test robot, RGB video streams were collected in a pedestrian walking environment. Continuous color images with a resolution of 640×640 pixels were extracted from the video stream. The collected images were manually annotated: a bounding rectangle was drawn around the entire human body on the image, semantic segmentation masks for the left hand, right hand, left leg, and right leg were applied, and a corresponding behavior label (one of five categories: standing, walking, running, falling, or bending over) was assigned to the current frame. Simultaneously, the pedestrian was continuously tracked, and the two-dimensional pixel coordinate sequence of the body's center point and key points of the limbs on the image was recorded over the next 5 seconds, constructing a multi-task behavior trajectory dataset.

[0035] Step 2: End-to-end training of the multi-task parallel network model

[0036] The labeled multi-task behavior trajectory dataset is randomly divided into training and validation sets in a 4:1 ratio. The following is a construction method: Figure 2 The multi-task parallel network model shown includes a shared backbone network, a multi-scale feature fusion network, and a multi-task processing branch. Training data is input into this network model. The network optimizer is set to AdamW, the initial learning rate is set to 0.001, the batch size is set to 16, and 200 iterations of training are performed. During training, a joint loss function is used to simultaneously correct the errors of the four outputs: bounding boxes, segmentation masks, behavior classification, and trajectory prediction, until the network converges. The trained model weight file is then saved.

[0037] The formula for the joint loss function is as follows:

[0038] ,

[0039] in, The preset weighting coefficients, For the human body localization bounding box loss, the CIoU loss function is specifically adopted. The intersection-union ratio, center point distance and aspect ratio similarity between the predicted bounding box and the real bounding box are calculated to constrain the localization accuracy of the overall spatial bounding box of the pedestrian. For limb semantic segmentation loss, a binary cross-entropy loss function is specifically adopted. This function applies a pixel-by-pixel category probability penalty to pixels in the image that belong to the human limb region, thereby constraining the boundary contour clarity of the limb semantic segmentation mask. For behavior classification loss, the Slide loss function is adopted. By introducing a dynamic hard sample adjustment coefficient based on the crossover ratio, the probability of judging five behavior categories (standing, walking, running, falling, and bending over) is weighted and constrained to solve the imbalance and missed detection problems of small sample behaviors during training. For dynamic trajectory prediction loss, the mean squared error loss function is specifically adopted. The mean squared difference between the pixel coordinates of the predicted limb key points and the actual key point pixel coordinates is calculated over N time steps within the next 5 seconds. This is used to constrain the coordinate accuracy of long-period trajectories in the two-dimensional image coordinate system.

[0040] Step 3: Real-time streaming sensing of video frame sequences

[0041] The trained model weight file is deployed to the embedded vehicle-mounted control chip of the mobile robot. During the robot's movement, the control chip continuously accesses the visible light camera mounted on the robot's front end, capturing the raw RGB video stream within the robot's field of view at a sampling frequency of F frames per second. The buffer within the control chip automatically extracts M consecutive frames of historical images from the past and the current moment, constructing a time-series image group.

[0042] Step 4: Multi-scale feature distribution in the backbone network

[0043] The time-series image sets are directly input into a shared backbone network, specifically an improved CSPDarknet backbone network. The input image size is 640×640 pixels, and the core processing architecture consists of 10 layers along the data flow: the first layer is a 2D convolutional layer with a kernel size of 3×3, a stride of 2, padding of 1, and 16 output feature map channels; the second layer is a 2D convolutional layer with a kernel size of 3×3, a stride of 2, padding of 1, and 32 output feature map channels; the third layer is a C3k2 feature extraction module with 32 output feature map channels, which contains a 2-layer cascaded lightweight bottleneck structure and uses the SiLU activation function; the fourth layer is a 2D convolutional layer with a kernel size of 3×3, a stride of 2, padding of 1, and 64 output feature map channels; the fifth layer is a C3k2... The module consists of three layers: a feature extraction module with 64 output feature map channels, containing two cascaded lightweight bottleneck structures and enabling residual shortcut connections; a 2D convolutional layer with a kernel size of 3×3, a stride of 2, and padding of 1, outputting 128 output feature map channels; a C3k2 feature extraction module with 128 output feature map channels, containing two cascaded lightweight bottleneck structures; a 2D convolutional layer with a kernel size of 3×3, a stride of 2, and padding of 1, outputting 256 output feature map channels; a C3k2 feature extraction module with 256 output feature map channels, containing two cascaded lightweight bottleneck structures; and a fast spatial pyramid pooling module containing three parallel 5×5 max pooling layers, with both input and output feature map channels fixed at 256.

[0044] The image group sequentially passes through the two-dimensional convolutional layer, C3k2 feature extraction module, and fast spatial pyramid pooling (SPPF) module of the backbone network. The 128-channel feature map output from the seventh layer, the 256-channel feature map output from the ninth layer, and the 256-channel feature map output from the tenth layer are output to the multi-scale feature fusion network. The multi-scale feature fusion network adopts a path aggregation network architecture with bidirectional fusion from top to bottom and bottom to top. The input end receives the feature maps distributed in parallel. Through cascaded upsampling and downsampling operations, the deep spatial geometric features and global semantic information are deeply fused to obtain deep spatial geometric feature vectors as basic features. The vectors are then distributed in parallel to the multi-task processing branches on demand: high-resolution feature maps rich in spatial topology and edge details are input to the limb segmentation branch to constrain the clarity of the mask boundary; low-resolution feature maps rich in macroscopic dynamic semantics are input to the behavior classification branch to improve the accuracy of intent determination; and multi-scale feature maps that take into account both positioning accuracy and motion trend are cascaded and input to the trajectory prediction branch as the initial latent feature matrix. This reduces the computational overhead of the embedded system while achieving efficient reuse of multi-task features.

[0045] Step 5: Parallel three-branch feature calculation and interaction

[0046] After the basic feature input is processed in the multi-task processing branch, the three branches perform the following operations simultaneously:

[0047] Limb segmentation branch: Utilizes a lightweight segmentation head to decode the input features, directly calculates and outputs the semantic segmentation mask of the pedestrian's limbs in the current image, completes the construction of the pedestrian's refined human topology in the image space dimension, and locks the pedestrian's fine spatial edges.

[0048] Behavior classification branch: The features of M consecutive frames are cascaded in time sequence using a temporal cascade encoder. The dynamic evolution features of pedestrian limbs on the time axis are analyzed through a temporal self-attention mechanism and compared with a preset temporal action primitive state library. Finally, the classification head outputs the specific probability values ​​of the current pedestrian belonging to one of the five types of actions: standing, walking, running, falling, and bending over. The preset behavior action types in the temporal action primitive state library include a total of five categories: standing state in which the body center remains basically still and the limbs do not move in a large range; walking state in which the limbs swing regularly and the walking speed is uniform and slow; running state in which the limbs swing with high amplitude and frequency and have motion acceleration and inertia; bending over state in which the upper body leans downward and the walking speed slows down sharply or stops in place; and falling state in which the body center moves violently downward and forward in a short period of time and then stops moving.

[0049] Trajectory Prediction Branch: During trajectory calculation, the behavior category probability distribution matrix output by the behavior classification branch is horizontally accessed. Let the behavior category probability vector output by the behavior classification branch be:

[0050] ,

[0051] Where C represents the total number of behavior categories; let the initial latent feature matrix of the trajectory prediction branch be... By constructing an attention mapping matrix The behavior classification probability is explicitly injected as a dynamic constraint into the trajectory features, resulting in a fused interaction feature matrix. Represented as:

[0052] ,

[0053] in, It represents the Hadamah accumulation. The Sigmoid activation function states that if the highest probability is for the running state, the inertia weight is automatically increased within the feature matrix; if the highest probability is for the bending or falling state, displacement constraints and speed drop weights are automatically applied to correct the motion trend within the feature space.

[0054] Step 6: Output of full limb motion trajectory and inverse coordinate projection

[0055] The multi-task parallel network model jointly outputs the results from the three branches. The decoder of the trajectory prediction branch, based on the corrected features, directly calculates and generates a sequence of absolute pixel coordinate predictions for the pedestrian's body center and limb key points in the image's two-dimensional coordinate system over the next 5 seconds (N time steps). Subsequently, the robot control chip reads the intrinsic and extrinsic parameter matrices of the visible light camera and the current real-time odometry data to establish a monocular geometric inverse projection model. Using the contact edge between the limb semantic segmentation mask output by the limb segmentation branch and the ground as a distance constraint, the two-dimensional image coordinate prediction trajectory is projected in real-time into a three-dimensional physical space trajectory coordinate sequence in the robot's world coordinate system.

[0056] Step 7: Cost Map Linkage Update and Active Obstacle Avoidance Execution

[0057] The control chip inputs the converted sequence of three-dimensional physical space trajectory coordinates for the next 5 seconds into the robot's navigation planner. The navigation planner dynamically draws an expanded safety protection boundary for each predicted time step within the next 5 seconds of the pedestrian's movement in a spatiotemporal cost map. If the pedestrian's current behavior is determined to be walking or bending over, the safety expansion radius around the future trajectory point is set as a baseline value. If the behavior is determined to be running or falling, the safety expansion radius is adjusted to... ,in By increasing the weight of collision-hazardous areas in the physical space, the local path planner is forced to switch to an alternative path in advance. The robot's local path planner reads the updated spatiotemporal dynamic cost map in real time, calculates the alternative or yielding path in advance, and sends control commands to the chassis drive motors to drive the mobile robot to complete proactive safety obstacle avoidance.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for active safety obstacle avoidance of robots based on human behavior understanding and spatiotemporal trajectory prediction, characterized in that, Includes the following steps: Step 1: Collect test RGB video stream and capture continuous color images. Jointly annotate the pedestrians in the images and add behavior labels. Record the two-dimensional pixel coordinate sequence of the pedestrian's body center point and key points of the limbs within 5 seconds to construct a multi-task behavior trajectory dataset. Step 2: Construct a multi-task parallel network model, including a shared backbone network, a multi-scale feature fusion network, and a multi-task processing branch. Train the model end-to-end using the multi-task behavior trajectory dataset. Step 3: Deploy the trained model, capture the original RGB video stream within the moving field of view in real time, extract historical images from the past and present of M consecutive frames, and construct a time-series image group; Step 4: Input the time series image group into the shared backbone network, and extract deep spatial geometric feature vectors as basic features through the multi-scale feature fusion network, and input them in parallel into the multi-task processing branches, including the limb segmentation branch, behavior classification branch and trajectory prediction branch; Step 5: Combine the calculation results of the multi-task processing branches to generate the predicted trajectory of the pedestrian's two-dimensional image coordinates within the next 5 seconds; Step 6: Convert the predicted trajectory of the two-dimensional image coordinates into a three-dimensional physical space trajectory coordinate sequence in the robot world coordinate system, and input it into the navigation planner. The navigation planner dynamically updates the safety protection boundary in the spatiotemporal dynamic cost map, and the local path planner reads and calculates the detour or deceleration path in real time to perform active safety obstacle avoidance.

2. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 1, characterized in that: The limb segmentation branch calculates and outputs the semantic segmentation mask of the pedestrian's limbs in the current image; the behavior classification branch concatenates the features of M consecutive frames in a time sequence, analyzes the limb swing pattern through a time self-attention mechanism, and outputs the behavior category determination probability of the current pedestrian; the trajectory prediction branch performs trajectory calculation, introduces the behavior category determination probability, and adaptively corrects the future trajectory movement trend.

3. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 1, characterized in that: The end-to-end training employs a joint loss function for collaborative optimization, as shown in the following formula: ， in, The preset weighting coefficients, For human body positioning frame loss, For limb semantic segmentation loss, Loss is categorized by behavior. Loss is calculated for dynamic trajectory prediction.

4. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 1, characterized in that: The shared backbone network is specifically an improved CSPDarknet backbone network. Its core processing architecture, along the data flow, consists of the following layers: a first 2D convolutional layer; a second 2D convolutional layer; a third C3k2 feature extraction module, which contains a cascaded lightweight bottleneck structure and uses the SiLU activation function; a fourth 2D convolutional layer; a fifth C3k2 feature extraction module, which contains a cascaded lightweight bottleneck structure and enables residual shortcut connections; a sixth 2D convolutional layer; a seventh C3k2 feature extraction module, which contains a cascaded lightweight bottleneck structure; an eighth 2D convolutional layer; a ninth C3k2 feature extraction module, which contains a cascaded lightweight bottleneck structure; and a tenth fast spatial pyramid pooling module, which contains a parallel max pooling layer.

5. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 4, characterized in that: During operation, the shared backbone network distributes the 128-channel feature map output from the seventh layer, the 256-channel feature map output from the ninth layer, and the 256-channel feature map output from the tenth layer in parallel to the multi-scale feature fusion network.

6. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 1, characterized in that: The multi-scale feature fusion network adopts a path aggregation network architecture with bidirectional fusion from top to bottom and bottom to top. The input end receives the feature map output by the shared backbone network. Through cascaded upsampling and downsampling operations, deep spatial geometric features and global semantic information are deeply fused. The feature map is then distributed in parallel to the multi-task processing branch on demand: high-resolution feature maps rich in spatial topology and edge details are input to the limb segmentation branch. Low-resolution feature maps rich in macroscopic dynamic semantics are input into the behavior classification branch; multi-scale feature maps that take into account both positioning accuracy and motion trend are cascaded and input into the trajectory prediction branch.

7. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 2, characterized in that: The trajectory prediction branch introduces a feature interaction mechanism that incorporates intent constraint weights, mathematically represented as follows: Let the behavior category determination probability vector output by the behavior classification branch be: ， in, Let be the total number of behavior categories; let the initial latent feature matrix of the trajectory prediction branch be... By constructing an attention mapping matrix The probability of behavior category determination is explicitly injected as a dynamic constraint into the trajectory features, resulting in a fused interaction feature matrix. Represented as: ， in, It represents the Hadamah accumulation. This represents the Sigmoid activation function.

8. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 2, characterized in that: The specific process of step 5 is as follows: The behavior classification branch extracts the limb swing amplitude and body center of gravity movement trend features of the pedestrian in the past M consecutive historical images through a temporal cascade encoder; the extracted features are matched with the standard action templates in the temporal action primitive state library to calculate the probability of the pedestrian's current behavior category; the decoder of the trajectory prediction branch adaptively calls the kinematic constraint law corresponding to the determined behavior action type to correct and generate the future trajectory.

9. The robot active safety obstacle avoidance method based on human behavior understanding and spatiotemporal trajectory prediction as described in claim 1, characterized in that: The linkage mechanism between behavior categories and safety expansion coefficients in the spatiotemporal dynamic cost map is as follows: if the pedestrian's current behavior is determined to be walking or bending over, then the safety expansion radius around the future trajectory point is set as the baseline value. ; If the behavior is determined to be running or falling, the safety expansion radius is adjusted to... ,in Increase the weight of collision hazard areas in physical space to force the local path planner to switch to the detour path in advance.

Images