Self-adaptive pedestrian motion trail prediction system based on laser ranging

Abstract

The invention relates to the technical field of pedestrian motion trail prediction, and discloses a self-adaptive pedestrian motion trail prediction system based on laser ranging. The system comprises a pedestrian image recognition and extraction module, a pedestrian posture image data determination module, a pedestrian motion trail prediction module and a pedestrian motion trail adjustment module. Firstly, a pedestrian motion image is acquired, multi-scale processing and an attention mechanism are fused to establish an initial pedestrian detection model, and a pedestrian contour is identified; secondly, pedestrian key points are extracted, attitude estimation is carried out, a pedestrian motion model is constructed, pedestrian motion image features are obtained, and an initial pedestrian motion prediction track is generated; and finally, introducing adaptive trajectory adjustment to adjust the initial pedestrian motion prediction trajectory, and considering environmental characteristics to obtain a final pedestrian motion prediction trajectory. According to the method, the pedestrian motion image is processed and modeled, the purpose of self-adaptive pedestrian motion track prediction is achieved, and the method is objective and accurate.

Description

Technical Field

[0001] This invention relates to the technical field of pedestrian trajectory prediction, specifically to an adaptive pedestrian trajectory prediction system based on laser ranging. Background Technology

[0002] In recent years, with the rapid development of autonomous driving, intelligent security, and robot navigation, pedestrian trajectory prediction has become one of the core technologies for ensuring system safety and interaction efficiency. Existing research mainly uses visual sensors, millimeter-wave radar, or multi-source data fusion to achieve pedestrian motion perception and performs trajectory prediction based on deep learning, probabilistic models, or social force models. However, pedestrian movement in complex urban environments is highly random, interactive, and environmentally dependent. Especially in scenarios with changing lighting, frequent occlusion, or dense crowds, existing methods still face significant challenges in terms of prediction accuracy, real-time performance, and dynamic adaptability, making it difficult to meet the needs of high-security applications.

[0003] Traditional methods for predicting pedestrian trajectories are typically based on kinematic models and statistical learning or computer vision and deep learning. For example, they may use convolutional neural networks (CNNs) to extract image features or long short-term memory networks (LSTMs) to learn temporal dependencies and fit pedestrian movement patterns to obtain pedestrian trajectories using historical trajectory data.

[0004] Traditional pedestrian trajectory prediction methods only consider historical trajectory data and lack an adaptive optimization mechanism. They cannot dynamically adjust the predicted trajectory based on the real-time pedestrian movement status (such as speed and direction of movement) and environmental characteristics, resulting in a conflict between the predicted trajectory and the physical environment and low accuracy of the predicted trajectory. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides an adaptive pedestrian trajectory prediction system based on laser ranging, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: This invention is an adaptive pedestrian trajectory prediction system based on laser ranging, specifically including: a pedestrian image recognition and extraction module, a pedestrian posture image data determination module, a pedestrian trajectory prediction module, and a pedestrian trajectory adjustment module; The pedestrian image recognition and extraction module is used to integrate multi-scale processing and attention mechanism to establish a pedestrian detection model and recognize pedestrian contours, and generate a set of processed pedestrian motion images. The pedestrian posture image data determination module is used to extract pedestrian key points from the processed pedestrian motion image set and estimate their posture to obtain pedestrian posture image data. The pedestrian trajectory prediction module is used to construct a pedestrian motion model based on pedestrian posture image data, obtain pedestrian motion image features, and generate an initial pedestrian motion prediction trajectory. The pedestrian trajectory adjustment module is used to adaptively adjust the initial pedestrian motion prediction trajectory based on laser ranging and pedestrian motion model to obtain the final pedestrian motion prediction trajectory.

[0007] Preferably, the pedestrian detection model established by fusing multi-scale processing and attention mechanisms includes: The process involves acquiring a set of pedestrian motion images and an initial set of pedestrian motion sample images, followed by data preprocessing and annotation to obtain the final set of pedestrian motion sample images. Using ResNet-50 as the encoder and transposed convolution and skip connections as the decoder, we added dilated spatial pyramid pooling to the encoder and introduced convolutional block attention modules at the skip connections to build the initial pedestrian detection model. The final pedestrian motion sample image set is divided into a sample image training set and a sample image test set. After normalizing the image pixels, the initial pedestrian detection model is trained using RGB three-channel input to obtain the final pedestrian detection model.

[0008] Preferably, the pedestrian contour recognition includes: The set of pedestrian motion images is input into the final pedestrian detection model, which outputs a pedestrian detection probability image. Then, it is subjected to binarization and morphological processing to output the pedestrian contour.

[0009] Preferably, the pedestrian key points extracted from the processed pedestrian motion image set include: Select a pedestrian motion reference image from the processed pedestrian motion image set, and mark the pedestrian torso key points, pedestrian lower limb key points, and pedestrian head key points to generate a pedestrian key point set.

[0010] Preferably, the attitude estimation includes: Establish a planar coordinate system, calculate the angle between the pedestrian's torso direction vector and the horizontal direction of the planar coordinate system, and obtain the pedestrian's body orientation; A three-dimensional coordinate system is established, and the knee flexion angle is calculated to obtain the change in pedestrian stride length; The pedestrian head deflection angle and the pedestrian head pitch angle are calculated based on the pedestrian key point set to form the pedestrian head direction; By combining the pedestrian's body orientation, stride length changes, and head direction, pedestrian posture image data is obtained.

[0011] Preferably, the step of constructing a pedestrian motion model based on pedestrian posture image data includes: Measure laser ranging data, establish a northeast-northeast coordinate system, and combine it with pedestrian posture image data to obtain pedestrian motion image data; Using the pedestrian's body orientation in the pedestrian motion image data as the pedestrian's initial motion direction, the correction amount for the pedestrian's motion direction is determined; A time interval threshold is set, and the pedestrian motion image data is compared with the time interval threshold to trigger a pedestrian acceleration mechanism, a pedestrian deceleration mechanism, and a pedestrian stopping mechanism, thereby establishing a pedestrian motion model.

[0012] Preferably, generating the initial pedestrian motion prediction trajectory includes: A safe distance threshold is set. When the laser ranging data in the pedestrian motion image data is less than the safe distance threshold, the pedestrian emergency response mechanism is triggered, and the emergency pedestrian walking speed is calculated. The pedestrian movement direction correction, pedestrian movement model, and emergency pedestrian walking speed are used as pedestrian motion image features. By comparing the deviation between the direction of the pedestrian's head and the direction of the pedestrian's body in the pedestrian motion image data, and then counting the number of times the pedestrian goes straight, turns left and turns right in the pedestrian motion image data, the probability of the pedestrian going straight, the probability of the pedestrian turning left and the probability of the pedestrian turning right are calculated. The pedestrian trajectory is predicted by comparing the probabilities of the pedestrian going straight, turning left, and turning right. The direction of movement with the highest probability is selected and the pedestrian moves again. Several pedestrian position coordinates are generated in sequence to obtain the initial predicted pedestrian trajectory.

[0013] Preferably, the adaptive adjustment of the initial pedestrian motion prediction trajectory based on laser ranging and pedestrian motion model includes: Obtain the pedestrian motion time series, determine the pedestrian motion model and emergency pedestrian speed of the initial pedestrian motion prediction trajectory, adaptively adjust the initial pedestrian motion prediction trajectory, and keep the laser ranging data not less than the safe distance threshold to obtain the processed pedestrian motion prediction trajectory. Environmental features are extracted, and the processed pedestrian motion prediction trajectory is controlled within the range of environmental features to obtain the final pedestrian motion prediction trajectory.

[0014] Preferably, the predicted pedestrian movement trajectory after control processing includes the following within the range of environmental features: Environmental features are extracted from the processed set of pedestrian motion images to ensure that the processed pedestrian motion prediction trajectory remains within the range of environmental features. Pedestrian motion prediction trajectories that exceed the range of environmental features are then removed to obtain the final pedestrian motion prediction trajectory.

[0015] The present invention has the following beneficial effects: 1. This invention acquires a set of pedestrian motion images, integrates multi-scale processing and attention mechanisms to establish an initial pedestrian detection model, identifies pedestrian contours, accurately detects and extracts pedestrians, and utilizes a combination of ResNet + dilated spatial pyramid pooling + DiceLoss to automatically learn the distinction between pedestrians and background using neural networks. Compared with other neural networks, this neural network has higher pedestrian contour recognition accuracy, and combined with morphological post-processing, it effectively improves mask quality, making it suitable for pedestrian motion analysis in complex scenes.

[0016] 2. This invention extracts key points of pedestrians and then estimates their posture, providing a high-precision spatial position basis for posture parameter calculation. It then determines pedestrian posture image data and uses multi-dimensional data integration to significantly avoid the limitations of static images. It makes full use of the pedestrian's motion state to achieve high-precision and real-time analysis of pedestrian posture, which helps to improve the accuracy of pedestrian motion trajectory prediction.

[0017] 3. This invention obtains pedestrian motion image features by constructing a pedestrian motion model, and uses the pedestrian motion direction correction amount to predict pedestrian turning in advance, which greatly improves the accuracy of pedestrian motion trajectory prediction. The constructed pedestrian motion model can identify more than 95% of sudden acceleration / sudden stop behaviors, generate initial pedestrian motion prediction trajectories, and can achieve trajectory prediction without relying on complex neural network models.

[0018] 4. This invention obtains pedestrian movement time series and considers environmental characteristics, introduces adaptive trajectory adjustment to adjust the predicted pedestrian movement trajectory, adjusts the movement trajectory in real time through the dynamic changes of pedestrians, and comprehensively considers the predicted trajectory and physical environment, sets hard boundaries, filters unreasonable trajectories through coordinate range, avoids pedestrian movement risks, and achieves high-precision prediction of pedestrian movement trajectory.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The present invention provides a flowchart of an adaptive pedestrian trajectory prediction system based on laser ranging.

[0022] Figure 2 This invention provides a flowchart of an adaptive pedestrian trajectory prediction method based on laser ranging. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Traditional pedestrian trajectory prediction methods only consider historical trajectory data and lack an adaptive optimization mechanism. They cannot dynamically adjust the predicted trajectory based on the real-time pedestrian movement status (such as speed and direction of movement) and environmental characteristics, resulting in a conflict between the predicted trajectory and the physical environment and low accuracy of the predicted trajectory.

[0025] To solve the above technical problems, such as Figure 1 As shown, this embodiment of the invention provides an adaptive pedestrian trajectory prediction system based on laser ranging, specifically including: a pedestrian image recognition and extraction module, a pedestrian posture image data determination module, a pedestrian trajectory prediction module, and a pedestrian trajectory adjustment module; the pedestrian image recognition and extraction module is used to acquire a set of pedestrian motion images, integrate multi-scale processing and attention mechanisms to establish an initial pedestrian detection model, identify pedestrian contours in the set of pedestrian motion images, and generate a processed set of pedestrian motion images; the pedestrian posture image data determination module is used to extract pedestrian key points in the processed set of pedestrian motion images, estimate pedestrian posture based on pedestrian key points, and obtain pedestrian posture image data; the pedestrian trajectory prediction module is used to combine laser ranging and pedestrian posture image data to obtain pedestrian motion image data, construct a pedestrian motion model using the pedestrian motion image data, obtain pedestrian motion image features, and generate an initial pedestrian motion prediction trajectory; the pedestrian trajectory adjustment module is used to acquire a pedestrian motion time series, introduce adaptive trajectory adjustment to adjust the initial pedestrian motion prediction trajectory, and consider environmental features to obtain a final pedestrian motion prediction trajectory.

[0026] In a specific embodiment, a camera is used to sample at 50Hz to collect pedestrian movement video, and each frame is used as a pedestrian movement image to provide good pedestrian image data for the pedestrian movement trajectory prediction of the present invention; then, vehicle-mounted radar is used to realize laser ranging, and the data obtained by laser ranging is considered to predict the pedestrian movement trajectory.

[0027] In the specific implementation process of the above embodiments, firstly, a set of pedestrian motion images is acquired. An initial pedestrian detection model is established by integrating multi-scale processing and attention mechanisms to identify pedestrian contours, accurately detect and extract pedestrians, and obtain a processed set of pedestrian motion images. This method uses a combination of ResNet + dilated spatial pyramid pooling + DiceLoss to automatically learn the distinction between pedestrians and background using a neural network. Compared with other neural networks, this neural network has higher accuracy in pedestrian contour recognition. Combined with morphological post-processing, it effectively improves mask quality and is suitable for pedestrian motion analysis in complex scenes. Secondly, pedestrian key points are extracted from the processed set of pedestrian motion images, and then the pedestrian's pose is estimated to determine the pedestrian's body orientation, stride length change, and head direction, obtaining pedestrian pose image data. This method simplifies pedestrian pose using pedestrian key points, providing a high-precision spatial position basis for pose parameter calculation. Combined with pedestrian body orientation, stride length change, and head direction, the multi-dimensional data integration significantly avoids the limitations of static images and fully utilizes the pedestrian's motion state. High-precision and real-time analysis of pedestrian posture helps improve the accuracy of pedestrian trajectory prediction. By combining laser ranging and pedestrian posture image data to generate pedestrian motion image data, a pedestrian motion model is constructed, pedestrian motion image features are obtained, and an initial pedestrian motion prediction trajectory is generated. Introducing laser ranging addresses the measurement error problem of a single sensor in rainy or foggy weather. Using pedestrian movement direction corrections to predict pedestrian turning in advance greatly improves the accuracy of pedestrian motion trajectory prediction. The constructed pedestrian motion model can identify more than 95% of sudden acceleration / sudden stop behaviors and can achieve trajectory prediction without relying on complex neural network models. Finally, by acquiring pedestrian motion time series and considering environmental characteristics, adaptive trajectory adjustment is introduced to adjust the initial pedestrian motion prediction trajectory, resulting in the final pedestrian motion prediction trajectory. This method calculates emergency walking speed in real time, considers dynamic changes in pedestrians, and can adjust the movement trajectory in real time. It comprehensively considers the predicted trajectory and the physical environment, using the zebra crossing area as a hard boundary and filtering unreasonable trajectories through coordinate range to avoid pedestrian movement risks and achieve pedestrian motion trajectory prediction.

[0028] Furthermore, to better illustrate the technical solution of the embodiments of the present invention, based on the above-mentioned adaptive pedestrian motion trajectory prediction system based on laser ranging, such as... Figure 2 As shown in the figure, this invention provides an adaptive pedestrian trajectory prediction method based on laser ranging, which specifically includes the following: S1. Acquire pedestrian motion images, form a pedestrian motion image set, integrate multi-scale processing and attention mechanism to establish an initial pedestrian detection model, identify pedestrian contours in the pedestrian motion image set, and generate a processed pedestrian motion image set. S1 includes the following steps: S11. Acquire pedestrian motion video, segment the pedestrian motion video by frame to obtain several pedestrian motion images, and form a pedestrian motion image set; obtain the COCO dataset from the network to obtain an initial pedestrian motion sample image set, randomly flip the initial pedestrian motion sample image set horizontally, and randomly crop it into sub-images of the same size to generate a final pedestrian motion sample image set; mark the pedestrian regions in the final pedestrian motion sample image set, marking the pedestrian regions as 1 and the image background as 0; S12. An initial pedestrian detection model is established by integrating multi-scale processing and attention mechanisms. The initial pedestrian detection model is then trained using the final set of pedestrian motion sample images to obtain the final pedestrian detection model. The specific steps are as follows: S121. Using ResNet-50 as the encoder, the decoder uses transposed convolutions and skip connections. Dilated spatial pyramid pooling is added to the encoder, and convolutional block attention modules are introduced at the skip connections. The loss function is DiceLoss, the optimizer is Adam, and the initial learning rate is... An initial pedestrian detection model is constructed; the final set of pedestrian motion sample images is divided into a sample image training set and a sample image test set, and the image pixels are normalized. S122. The sample image training set is used to train the initial pedestrian detection model using RGB three-channel input. ResNet-50 extracts features and outputs feature maps. Transposed convolution is used for progressive upsampling. A maximum number of training epochs is set, and the initial learning rate decays by 1% each time until the maximum number of training epochs is reached, at which point training stops, resulting in a trained pedestrian detection model. The sample image test set is then input into the trained pedestrian detection model, and pedestrian detection results are output. A precision threshold is set. When the precision of the pedestrian detection result is greater than the precision threshold, the final pedestrian detection model is obtained. Otherwise, sample images with pedestrian detection results whose precision is not greater than the precision threshold are selected and repeated for training until the precision of the pedestrian detection result is greater than the precision threshold. S13. Input the set of pedestrian motion images into the final pedestrian detection model, output the pedestrian detection probability image, perform binarization processing on the pedestrian detection probability image to generate a pedestrian motion image mask, perform dilation and erosion processing on the pedestrian motion image mask, and then perform smoothing processing to output the pedestrian contour and obtain the processed set of pedestrian motion images. In this embodiment, a set of pedestrian motion images is acquired, and an initial pedestrian detection model is established by fusing multi-scale processing and attention mechanisms. Pedestrian contours are identified, pedestrians are accurately detected and extracted, resulting in a processed set of pedestrian motion images. This method utilizes a combination of ResNet, dilated spatial pyramid pooling, and DiceLoss to automatically learn the distinction between pedestrians and the background. Compared to other neural networks, this neural network offers higher accuracy in pedestrian contour recognition. Combined with morphological post-processing, it effectively improves mask quality and is suitable for pedestrian motion analysis in complex scenes. Specifically, for example, the encoder uses a ResNet-50 network to extract image features and introduces a dilated spatial pyramid pooling (ASPP) module during the encoding stage to fuse multi-scale contextual information. The decoder uses transposed convolution to progressively upsample and restore image resolution, combining high-level features from the encoder with low-level features from the decoder through skip connections. Additional features are added at the skip connections. The Convolutional Block Attention (CBAM) module enhances the ability to focus on the spatial and channel features of pedestrian regions. Features are extracted by ResNet-50 and multi-scale feature maps are output. After processing by the ASPP and CBAM modules, a pedestrian probability map is generated by the decoder. DiceLoss is used to optimize segmentation accuracy. The initial learning rate is set to 0.001, decaying by 1% per training epoch, with a maximum of 100 iterations. The final sample set is divided into a training set (6400 images) and a test set (1600 images) in an 8:2 ratio. Image pixel values ​​are normalized to the range [0,1]. If the accuracy of the test set does not reach the 95% accuracy threshold, low-precision samples are added to the training set for repeated training until the threshold is met. The mask's internal holes are filled and broken areas are connected through dilation, and small noise points are eliminated through erosion. The processed mask is then superimposed on the original image to obtain an image set containing accurate pedestrian contours for subsequent behavior analysis or tracking tasks. S2. Extract pedestrian key points from the processed pedestrian motion image set, estimate pedestrian posture based on pedestrian key points, determine pedestrian body orientation, pedestrian stride length change and pedestrian head direction, and obtain pedestrian posture image data. S2 includes the following steps: S21. Select a pedestrian motion image from the processed pedestrian motion image set and denote it as a pedestrian motion reference image. Select key points of the pedestrian torso from the pedestrian motion reference image. The key points of the pedestrian torso include the two shoulder points, two hip points, and the midpoint of the spine. Select key points of the pedestrian lower limbs from the pedestrian motion reference image. The key points of the pedestrian lower limbs include the hip point, knee point, ankle point, and heel point. Select key points of the pedestrian head from the pedestrian motion reference image. The key points of the pedestrian head include the left ear point, right ear point, left eye point, and right eye point. Combine the key points of the pedestrian torso, lower limbs, and head to obtain pedestrian key points, and label them to generate a pedestrian key point set. S22. Based on pedestrian key points, estimate pedestrian posture to determine pedestrian body orientation, stride length changes, and head direction, thereby obtaining pedestrian posture image data. The specific steps are as follows: S221. Select key points of the pedestrian torso from the set of key points of the pedestrian torso, connect the two shoulder points and two hip points among the key points of the pedestrian torso to form a planar quadrilateral, establish a planar coordinate system on the planar quadrilateral, obtain the coordinates of the two shoulder points and the two hip points, and establish a coordinate matrix of size 4×2. A Calculate the covariance matrix Furthermore, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain several eigenvalues. The eigenvector corresponding to the largest eigenvalue is selected and denoted as the pedestrian torso direction vector. The angle between the pedestrian torso direction vector and the horizontal direction of the plane coordinate system is calculated to obtain the pedestrian's body orientation. S222. Select key points of the lower limbs of the pedestrian from the set of key points of the pedestrian. Take the midpoint of the hip as the origin of the coordinate system, the direction of the pedestrian's body as the X-axis, and the vertical direction as the Z-axis to establish a three-dimensional coordinate system; connect the hip point, the knee point, and the ankle point to calculate the knee flexion angle. ,in This represents the vector pointing from the hip point to the knee point. The vector representing the point from the ankle to the knee. Obtain adjacent frame images of the pedestrian motion reference image to obtain a pedestrian motion reference image set. When the knee joint flexion angle changes, calculate the change in pedestrian stride length based on the pedestrian motion reference image set. The specific steps are as follows: S2221. Set a height threshold. When the heel height of the continuous pedestrian motion reference image in the pedestrian motion reference image set is greater than the height threshold, obtain the touch point coordinates. Calculate the Euclidean distance and time interval of the continuous touch point coordinates in the continuous pedestrian motion reference images to obtain the pedestrian stride length and pedestrian gait cycle, respectively. S2222. Calculate the total time and total number of steps in the set of pedestrian motion reference images to obtain the pedestrian gait frequency. Then calculate the product of the pedestrian stride length and the pedestrian gait frequency to obtain the pedestrian walking speed. Combine the pedestrian gait period and the pedestrian walking speed to form the pedestrian stride length variation. S223. Select pedestrian head key points from the set of pedestrian key points. When the left ear point is to the left of the left eye point and the right ear point is to the right of the right eye point, the pedestrian's head is facing forward, and the head deflection angle is 0. When the left ear point and the right eye point overlap, the pedestrian's head is turning right. When the right ear point and the left eye point overlap, the pedestrian's head is turning left. ,in This indicates the horizontal displacement of the left ear point and the left eye point. The distance between the left and right eye points is represented; the left ear point, right ear point, and nose tip point are selected from the key points of the human head to construct the head reference plane, and the coordinate system of the head reference plane is obtained. At this point, the normal vector of the head reference plane is calculated. ,in Indicates the coordinates of the left ear point. Indicates the coordinates of the right ear point. This represents the coordinates of the tip of the nose, at which point the pedestrian's head tilts. ,in Represents the normal vector of the head reference plane z Coordinate values; the pedestrian head tilt angle and head pitch angle are taken as the pedestrian head direction; S224. By combining the pedestrian's body orientation, walking speed, gait cycle, and head direction, pedestrian posture image data is obtained. In this embodiment, pedestrian key points are extracted from the processed pedestrian motion image set, and then the pedestrian's posture is estimated to determine the pedestrian's body orientation, stride length change, and head direction, thus obtaining pedestrian posture image data. This method simplifies pedestrian posture by using pedestrian key points, providing a high-precision spatial position basis for posture parameter calculation. Combined with pedestrian body orientation, stride length change, and head direction, the multi-dimensional data integration significantly avoids the limitations of static images, fully utilizing the pedestrian's motion state to achieve high-precision and real-time analysis of pedestrian posture. Specifically, for example, key points for the torso are set as follows: left shoulder (10, 20), right shoulder (30, 20), left hip (...). Key points of the lower limbs: left hip (15, 50), right hip (25, 50), midpoint of the spine (20, 35); right hip (25, 50), right knee (15, 70), left ankle (15, 90), left heel (10, 90); right hip (25, 50), right knee (25, 70), right ankle (25, 90), right heel (30, 90); key points of the head: left ear (5, 10), right ear (35, 10), left eye (15, 15), right eye (25, 15), tip of nose (20, 10); pedestrian body orientation calculation: construct coordinate matrix, after feature decomposition, the eigenvector corresponding to the largest eigenvalue is (0.8, 0.6), the included angle is calculated to be 36.9°. The pedestrian's body is facing right front, with an angle of approximately 36.9° to the horizontal. Analysis of pedestrian stride length changes: The knee flexion angle is calculated to be 180°, which is the gait support phase. In five consecutive frames, the right heel height decreases from 95 pixels to 85 pixels (the height threshold is determined by the pedestrian's walking habits in the motion image and set to 90 pixels). The Euclidean distance between adjacent contact points is 30 pixels (0.5 seconds interval), the stride length is 30 pixels, the gait cycle is 0.5 seconds, and 10 steps are detected within a total duration of 5 seconds. The stride frequency is 2 steps / second, and the stride speed = stride length × stride frequency = 30 × 2 = 60 pixels / second. Therefore, the pedestrian gait cycle is 0.5 seconds, and the stride speed is 60 pixels / second. Pedestrian head direction determination: Left ear (5, 10) With the left eye (15, 15) to the left (horizontal coordinate 5 < 15) and the right ear (35, 10) to the right of the right eye (25, 15) (35 > 25), determine the head orientation. When the head rotates, assuming the right ear (25, 15) and the left eye (15, 15) are horizontally overlapping, the horizontal displacement is 10, the inter-eye distance is 10, and the pedestrian head deflection angle is 45°. The calculated pedestrian head pitch angle is 30°. The pedestrian head deflection angle is set between 45 degrees to the right and 45 degrees to the left, and the pedestrian head pitch angle is between -90 degrees and +60 degrees. The entire process, through key point structured calibration, multi-dimensional parameter fusion, and dynamic spatiotemporal analysis, helps to improve the accuracy of pedestrian motion trajectory prediction. S3. Based on laser ranging and combined with the pedestrian posture image data, obtain pedestrian motion image data, use the pedestrian motion image data to construct a pedestrian motion model, obtain pedestrian motion image features, and generate an initial pedestrian motion prediction trajectory; S3 includes the following steps: S31. Select the target vehicle and use a laser rangefinder to measure the distance from the target vehicle to the pedestrian, obtaining the laser ranging data; at this point, establish a northeast-central coordinate system with the target vehicle as the origin to obtain the pedestrian's initial position coordinates. Combined with pedestrian posture image data, the pedestrian's body orientation pedestrian walking speed v pedestrian gait cycle and the direction of pedestrian's head Obtain pedestrian motion image data ,in d This represents laser ranging data; S32. Using the pedestrian motion image data, establish a pedestrian motion model to obtain pedestrian motion image features. The specific steps are as follows: S321. Using the pedestrian's body orientation in the pedestrian motion image data as the pedestrian's initial movement direction, a rotation time threshold is set. The pedestrian motion image corresponding to the change in pedestrian head orientation in the processed pedestrian motion image set is identified and denoted as the pedestrian head orientation change image. The number of times the pedestrian's body orientation changes after the rotation time threshold in the pedestrian head orientation change image is counted, the probability of the pedestrian's body orientation change is calculated, and then the pedestrian movement direction correction amount is calculated. ,in c This indicates the probability of a pedestrian changing their body orientation; S322. Set a time interval threshold. When the pedestrian gait cycle in the pedestrian motion image data continuously decreases within the time interval threshold, trigger a pedestrian acceleration mechanism; when the pedestrian gait cycle in the pedestrian motion image data continuously increases within the time interval threshold, trigger a pedestrian deceleration mechanism; when the pedestrian gait cycle in the pedestrian motion image data is 0 within the time interval threshold, trigger a pedestrian stopping mechanism and establish a pedestrian motion model. S323. Set a safe distance threshold. When the laser ranging data in the pedestrian motion image data is less than the safe distance threshold, trigger the pedestrian emergency response mechanism. At this time, the pedestrian moves at an emergency pedestrian pace, and the emergency pedestrian pace is calculated. The pedestrian movement direction correction, pedestrian movement model, and emergency pedestrian walking speed are used as pedestrian movement image features. S33. Compare the deviation between the pedestrian's head direction and the pedestrian's body direction in the pedestrian motion image data, set a deviation threshold, and when the deviation between the pedestrian's head direction and the pedestrian's body direction is greater than the deviation threshold and the pedestrian's head pitch angle is less than the deviation threshold, the pedestrian begins to turn. Count the number of times the pedestrian moves straight, turns left and turns right in the pedestrian motion image data, and calculate the probability of the pedestrian moving straight, the probability of the pedestrian turning left and the probability of the pedestrian turning right. Given the pedestrian's initial position coordinates, predicting the pedestrian's trajectory will yield the following next position coordinates: ,in , At the next position coordinate of the pedestrian, compare the probability of the pedestrian going straight, the probability of the pedestrian turning left and the probability of the pedestrian turning right, select the highest probability as the direction of the pedestrian's movement and move again, generate several pedestrian position coordinates in sequence, and connect several pedestrian position coordinates in sequence to obtain the initial pedestrian movement prediction trajectory. In this embodiment, pedestrian motion image data is generated by combining laser ranging and pedestrian posture image data, a pedestrian motion model is constructed, pedestrian motion image features are obtained, and an initial pedestrian motion prediction trajectory is generated. By introducing laser ranging, the measurement error problem of a single sensor in rainy or foggy weather is solved. Pedestrian movement direction correction is used to predict pedestrian turning in advance, greatly improving the accuracy of pedestrian motion prediction trajectory. The constructed pedestrian motion model can identify more than 95% of sudden acceleration / sudden stop behaviors and can achieve trajectory prediction without relying on complex neural network models. Specifically, for example, the initial position coordinates of the pedestrian are set to (0, 8), the pedestrian motion feature parameter set is [0, 8, 90°, 1.2, 0.9, 60°, 8], the rotation time threshold is 2 seconds, and if the pedestrian's head direction is detected to continuously turn from 60° to 30° within 2 seconds, based on the image... The data calculates a body orientation change probability of 0.4, correcting the movement direction by 54°; a time interval threshold of 3 seconds is set, decreasing by 0.2 seconds within 3 seconds triggers an acceleration mechanism, updating the walking speed to 1.32 m / s; a safe distance threshold of 5 meters is set, currently 8 meters > 5 meters, so no emergency response is triggered; a turning deviation threshold of 15° is set, currently the head direction deviates from the body orientation by 30° > 15°, and the pedestrian's head pitch angle is 5° < 15°. Historical data is collected: 6 straight steps, 2 left turns, and 1 right turn. The calculated probability of the pedestrian going straight is 67%, turning left is 22%, and turning right is 11%. Therefore, the first step predicts (0.53, 8.75), generating 7 predicted coordinate points for the next 5 seconds. Connecting these coordinate points forms a predicted trajectory shifting northeast; the error between the initial predicted pedestrian trajectory and the actual pedestrian trajectory is 18%. S4. Obtain the pedestrian movement time series, introduce adaptive trajectory adjustment to adjust the initial pedestrian movement prediction trajectory, and consider environmental characteristics to obtain the final pedestrian movement prediction trajectory, thus completing the pedestrian movement trajectory prediction. S4 includes the following steps: S41. After generating the initial pedestrian motion prediction trajectory, obtain the pedestrian motion time series, and adjust the initial pedestrian motion prediction trajectory based on the pedestrian motion image features using adaptive trajectory adjustment. The specific steps are as follows: S411. Traverse all pedestrian position coordinates in the initial pedestrian motion prediction trajectory, select any pedestrian position coordinate as the pedestrian position to be adjusted; collect pedestrian motion time points to obtain pedestrian motion time series, find the pedestrian motion time point corresponding to the pedestrian position to be adjusted; consider the pedestrian motion model in the pedestrian motion image features, determine the pedestrian motion model at the pedestrian position to be adjusted, and sequentially determine the pedestrian motion model of the initial pedestrian motion prediction trajectory; S412. Calculate the emergency pedestrian speed at the pedestrian position to be adjusted, determine the emergency pedestrian speed of the initial pedestrian movement prediction trajectory according to the pedestrian movement time sequence, adaptively adjust the initial pedestrian movement prediction trajectory according to the pedestrian movement model of the initial pedestrian movement prediction trajectory, and keep the laser ranging data not less than the safe distance threshold to obtain the processed pedestrian movement prediction trajectory. S42. Extract environmental features from the processed pedestrian motion image set. The environmental features include sidewalks and zebra crossings. Ensure that the processed pedestrian motion prediction trajectory moves within the range of environmental features. Remove pedestrian motion prediction trajectories that exceed the range of environmental features to obtain the final pedestrian motion prediction trajectory and complete the pedestrian motion trajectory prediction. In this embodiment, a pedestrian movement time series is acquired, and considering environmental characteristics, an adaptive trajectory adjustment is introduced to adjust the initial pedestrian movement prediction trajectory to obtain the final pedestrian movement prediction trajectory. This method calculates the emergency walking speed in real time, considers the dynamic changes of pedestrians, and can adjust the movement trajectory in real time. It also comprehensively considers the predicted trajectory and the physical environment, using the zebra crossing area as a hard boundary, and filters unreasonable trajectories through coordinate range to avoid pedestrian movement risks and achieve pedestrian movement trajectory prediction. Specifically, for example, the data is collected at time points, each time point corresponding to a position coordinate. When t=3 seconds, a vehicle on the left is detected to be only 20 meters away from the pedestrian and approaching at a speed of 10 m / s. The relative speed between the vehicle and the pedestrian (the vehicle needs 3 seconds to reach the pedestrian's position) is calculated. The system determines that pedestrians need to accelerate to 2.5 m / s within 2 seconds to cross the danger zone in advance. The pedestrian motion model switches from uniform speed to acceleration, and the trajectory coordinates are adjusted after t≥3 seconds. The new trajectory position at t=3 seconds is (4.5, 0), and the position is updated to (4.5+2.5, 0), (7+2.5, 0), etc. every second thereafter, while maintaining the minimum distance between the adjusted trajectory and the vehicle at least 5 meters. The zebra crossing range is set as x∈[0, 15], y∈[-1, 1] (width 2 meters), and the sidewalk range is y≥15 (sidewalk starting point). If y=1.5 after adjustment (exceeding the zebra crossing), the trajectory is discarded and readjusted. The error between the adjusted trajectory and the actual pedestrian trajectory is 13%, achieving high-precision prediction of pedestrian motion trajectory.

[0029] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0030] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. An adaptive pedestrian trajectory prediction system based on laser ranging, characterized in that, include: The pedestrian image recognition and extraction module is used to integrate multi-scale processing and attention mechanisms to establish a pedestrian detection model and identify pedestrian contours, generating a set of processed pedestrian motion images. The pedestrian pose image data determination module is used to extract pedestrian key points from the processed pedestrian motion image set and estimate the pose to obtain pedestrian pose image data. The pedestrian trajectory prediction module is used to construct a pedestrian motion model based on pedestrian posture image data, obtain pedestrian motion image features, and generate an initial pedestrian motion prediction trajectory. The pedestrian trajectory adjustment module is used to adaptively adjust the initial pedestrian motion prediction trajectory based on laser ranging and pedestrian motion model to obtain the final pedestrian motion prediction trajectory.

2. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 1 The system is characterized by, The pedestrian detection model, which integrates multi-scale processing and attention mechanisms, includes: The process involves acquiring a set of pedestrian motion images and an initial set of pedestrian motion sample images, followed by data preprocessing and annotation to obtain the final set of pedestrian motion sample images. Using ResNet-50 as the encoder and transposed convolution and skip connections as the decoder, we added dilated spatial pyramid pooling to the encoder and introduced convolutional block attention modules at the skip connections to build the initial pedestrian detection model. The final pedestrian motion sample image set is divided into a sample image training set and a sample image test set. After normalizing the image pixels, the initial pedestrian detection model is trained using RGB three-channel input to obtain the final pedestrian detection model.

3. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 2 The system is characterized by, The pedestrian contour recognition includes: The set of pedestrian motion images is input into the final pedestrian detection model, which outputs a pedestrian detection probability image. Then, it is subjected to binarization and morphological processing to output the pedestrian contour.

4. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 3 The system is characterized by, The extracted and processed pedestrian motion image set includes the following pedestrian key points: Select a pedestrian motion reference image from the processed pedestrian motion image set, and mark the pedestrian torso key points, pedestrian lower limb key points, and pedestrian head key points to generate a pedestrian key point set.

5. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 4 The system is characterized by, The attitude estimation includes: Establish a planar coordinate system, calculate the angle between the pedestrian's torso direction vector and the horizontal direction of the planar coordinate system, and obtain the pedestrian's body orientation; A three-dimensional coordinate system is established, and the knee flexion angle is calculated to obtain the change in pedestrian stride length; The pedestrian head deflection angle and the pedestrian head pitch angle are calculated based on the pedestrian key point set to form the pedestrian head direction; By combining the pedestrian's body orientation, stride length changes, and head direction, pedestrian posture image data is obtained.

6. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 5 The system is characterized by, The step of constructing a pedestrian motion model based on pedestrian posture image data includes: Measure laser ranging data, establish a northeast-northeast coordinate system, and combine it with pedestrian posture image data to obtain pedestrian motion image data; Using the pedestrian's body orientation in the pedestrian motion image data as the pedestrian's initial motion direction, the correction amount for the pedestrian's motion direction is determined; A time interval threshold is set, and the pedestrian motion image data is compared with the time interval threshold to trigger a pedestrian acceleration mechanism, a pedestrian deceleration mechanism, and a pedestrian stopping mechanism, thereby establishing a pedestrian motion model.

7. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 6 The system is characterized by, The generation of the initial pedestrian motion prediction trajectory includes: A safe distance threshold is set. When the laser ranging data in the pedestrian motion image data is less than the safe distance threshold, the pedestrian emergency response mechanism is triggered, and the emergency pedestrian walking speed is calculated. The pedestrian movement direction correction, pedestrian movement model, and emergency pedestrian walking speed are used as pedestrian motion image features. By comparing the deviation between the direction of the pedestrian's head and the direction of the pedestrian's body in the pedestrian motion image data, and then counting the number of times the pedestrian goes straight, turns left and turns right in the pedestrian motion image data, the probability of the pedestrian going straight, the probability of the pedestrian turning left and the probability of the pedestrian turning right are calculated. The pedestrian movement trajectory is predicted by comparing the probabilities of the pedestrian going straight, turning left, and turning right. The direction of movement with the highest probability is selected and the pedestrian moves again. Several pedestrian position coordinates are generated in sequence to obtain the initial pedestrian movement prediction trajectory.

8. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 7 The system is characterized by, The adaptive adjustment of the initial pedestrian motion prediction trajectory based on laser ranging and pedestrian motion model includes: Obtain the pedestrian motion time series, determine the pedestrian motion model and emergency pedestrian speed of the initial pedestrian motion prediction trajectory, adaptively adjust the initial pedestrian motion prediction trajectory, and keep the laser ranging data not less than the safe distance threshold to obtain the processed pedestrian motion prediction trajectory. Environmental features are extracted, and the processed pedestrian motion prediction trajectory is controlled within the range of environmental features to obtain the final pedestrian motion prediction trajectory.

9. The adaptive pedestrian trajectory prediction system based on laser ranging according to claim 8 The system is characterized by, The predicted pedestrian movement trajectory after control processing includes, within the range of environmental features: Environmental features are extracted from the processed set of pedestrian motion images to ensure that the processed pedestrian motion prediction trajectory moves within the range of environmental features. Pedestrian motion prediction trajectories that exceed the range of environmental features are removed to obtain the final pedestrian motion prediction trajectory.

10. An adaptive pedestrian trajectory prediction method based on laser ranging as described in any one of claims 1-9, characterized in that, Specifically, it includes: S1. Acquire pedestrian motion images, form a pedestrian motion image set, integrate multi-scale processing and attention mechanism to establish an initial pedestrian detection model, identify pedestrian contours in the pedestrian motion image set, and generate a processed pedestrian motion image set. S2. Extract pedestrian key points from the processed pedestrian motion image set, estimate pedestrian posture based on pedestrian key points, determine pedestrian body orientation, pedestrian stride length change and pedestrian head direction, and obtain pedestrian posture image data. S3. Based on laser ranging and combined with the pedestrian posture image data, obtain pedestrian motion image data, use the pedestrian motion image data to construct a pedestrian motion model, obtain pedestrian motion image features, and generate an initial pedestrian motion prediction trajectory; S4. Obtain the pedestrian movement time series, introduce adaptive trajectory adjustment to adjust the initial pedestrian movement prediction trajectory, and consider environmental characteristics to obtain the final pedestrian movement prediction trajectory, thus completing the pedestrian movement trajectory prediction.

Images