Research On Spatial And Temporal Behavior Prediction Algorithm Of The Surrounding Environment Of Urban Intelligent Vehicle

When an urban intelligent vehicle is driving,it will inevitably interact with surrounding traffic participants.Predicting the behavior of the surrounding traffic participants is a key ability to ensure the safe and stable driving of the urban intelligent vehicle.based on the observation information of perceived objects to predict the trajectory and behavior of objects in the future.Therefore,the research on the prediction algorithm of the spatio-temporal behavior of the objects around the urban intelligent vehicle mainly includes the following contents:Aiming at the problem that the prediction algorithm depends on the input of perceptual information,this paper designs 3D object detection algorithm and tracking algorithm to obtain high-precision perception information of objects.Based on the characteristics of image and lidar point cloud,this paper proposes to fuse image and lidar point cloud to realize 3D object detection.By analyzing the advantages and disadvantages of several current fusion detection algorithms,this paper designs a multi-view-pointcloud fusion dynamic object 3D detection algorithm.The algorithm use image detection algorithm and bird's-eye-view(BEV)detection algorithm YOLO-BEV which designed based on image detection network,to detect objects from both image and BEV to make up for the missing objects from image.In order to improve the box regression accuracy,the algorithm use the points in 2d detection box to estimate the 3D box parameters of objects.The accuracy and max-recall rate of the detection algorithm are verified on dataset.Based on the detection results,this paper designs a multi-target 3D tracking algorithm,by constructing a 3D Kalman filter and a Hungarian matching algorithm to track objects in real time.The experiment verifies that the tracking algorithm in this paper improves the ID switch in Sort.Aiming at the object trajectory prediction problem,by analyzing the characteristics of the object motion in the urban scene affected by other surrounding objects and the surrounding static scenes,this paper designs a dynamic object trajectory prediction algorithm based on interaction and map features,the input is perceived object for 2 seconds observation trajectory and bird's eye view map.By constructing a spatiotemporal map,an interactive feature extraction network based on interactive attention mechanism is used to extract the interactive features from the observation trajectory to capture the pairwise interaction effects between objects.A map feature extraction network based on map attention mechanism is designed to extract the map feature to capture the effects of map on object motion.Finally,the interactive features and map features are fused to predict,and then output the predicted position of the object in the next 3 seconds.In order to verify the performance of the algorithm,this paper produced a dataset for network training and comparison experiments.The experiment verifies that considering both the interaction of objects and the map features can reduce the prediction error.Aiming at the object behavior intention prediction problem,by analyzing the characteristics of the object behavior intention information implied by the prediction trajectory and different types of objects with different behavior patterns,this paper designs different behavior recognition algorithms for different objects based on prediction trajectory.Finally,this paper tests the above algorithms on the urban autonomous driving scene dataset in turn to verify the effectiveness of the algorithm.In order to apply the research in this paper to ensure the driving safety of urban intelligent vehicle,this paper designs a collision warning algorithm based on the above research results to provide collision warning information of surrounding objects for urban intelligent vehicle,so that it can be safely and reasonably planned.Then build the overall algorithm on the real autonomous driving vehicle to experiment on campus.In summary,aiming at the behavior prediction of objects around urban intelligent vehicles,this paper proposes a multi-view-point-cloud fusion dynamic object 3D detection and tracking algorithm,a dynamic object trajectory prediction algorithm based on interaction and map features,and a behavior recognition algorithm based on predicted trajectory.Finally,the effectiveness of the proposed algorithm is verified on urban intelligent driving scene dataset and real autonomous driving vehicle,and a collision warning algorithm is designed to apply this research to ensure the driving safety of urban intelligent vehicles.