Research And Design Of Multi-Object Tracking Method For Vehicle-Road Coordination

The increase in the number of cars has brought convenience to people’s daily travel.The emergence of intelligent transportation systems has further improved travel efficiency.As part of the intelligent transportation system,multi-object tracking can provide a basis for tasks such as path planning and traffic behavior analysis and improve the network efficiency and security of road traffic.The traditional object tracking model is based on on-board sensors to complete the multi-object tracking task.The deployment position of the on-board sensors is low and easy to be blocked,and moves with the vehicle,which affects the accuracy of multi-object tracking.Some recent researches fuse vehicle-mounted sensors and roadside sensors during tracking,and multi-source data fusion takes a long time,which affects the real-time performance of multi-object tracking.This thesis proposes a multi-object tracking method for vehicle-road collaboration,which integrates vehicle-road data in real time,and makes full use of temporal and spatial information to improve the accuracy of multi-object tracking.Considering the temporal and spatial cues of multi-object tracking tasks,this thesis proposes a multi-object tracking model based on spatiotemporal features.This model fully considers the spatio-temporal features of the multi-object tracking task,designs a graph convolution model,fully aggregates the temporal and spatial features between target objects,and obtains more representative target features.This model largely alleviates the problem of association failure due to occlusion or camera motion,which improves 1.7%and 4%HOTA and MOTA performance on the KITTI dataset vehicle category data respectively.Considering the over-the-horizon perception ability of roadside sensors,and the high deployment position,which is not easily affected by occlusion,this thesis proposes a multi-object tracking model for vehicleroad coordination.This model fuses the vehicle and road data in real time through a post-fusion method and expands the graph convolution model at the same time.This method integrates multi-source data to realize overthe-horizon perception,alleviates the problem of association failure caused by long-term occlusion and deformation,and improves the accuracy of multi-target tracking.Experiments show that the multi-object tracking model for vehicle-road collaboration improves the performance of HOTA and MOTA by 1.1%and 2.5%on the Carla dataset vehicle category data respectively.