| Autonomous driving has gradually become an important content in the field of intelligent transportation.Object tracking is one of its key technologies,and there are two mainstream paradigms: end-to-end(E2E)and detection-to-tracking(D2T).The former is leading in terms of speed,but the accuracy cannot meet industry needs? the latter uses detection and tracking for inter-frame matching of detection and trajectory,and is more close to the accuracy requirements of traffic scenes.However,object tracking still faces many challenges: in the detection stage,the performance of most industrial equipment is difficult to support actual deployment,and the detection effect of small object is poor?in the tracking stage,the dynamic change of objects in the video stream makes tracking difficult.Therefore,accuracy and speed are difficult to achieve a balance.Based on this,this paper takes two-dimensional moving objects as the research object,and combines deep learning,object detection,projection,and clustering methods to study the online object tracking of the D2 T paradigm.Because the detection stage of the D2 T paradigm has a great impact on the subsequent tracking accuracy,a progressive research approach is adopted,starting from feature extraction leading to detection,then from detection leading to target tracking,and finally studying the entire tracking stage.All three contents consider the lightness of the model and the robustness of the scale,thus achieving target tracking with high efficiency.This research has certain significance for intelligent transportation and autonomous driving.The main contents of this thesis include:(1)Proposing a lightweight feature extractor PG-PRNet based on parallel gate units.In order to improve the feature expression ability of the lightweight network,a threshold switch is added to the parallel gate unit branch part to make the branch have the ability to adjust information output according to the step size.In order to alleviate the unreasonable problem of using the same regularization rate for different scale feature maps and network sizes,an adaptive progressive regularization algorithm is proposed to allow the model to automatically adjust the regularization factor during training.Finally,channel pruning,Hard Sigmoid,Depth-wise convolution and random depth methods are used to further reduce the model calculation.Through ablation and visualization experiments,it is proved that PG-PRNet has good feature extraction ability in image recognition tasks.(2)Constructing the object detector V-PGPR based on PG-PRNet and YOLOV4.To alleviate the influence of dynamic changes in the object area of the video stream on the model input,the baseline model is modified with variable input,and convolution is used instead of full connection layer to achieve local feature response,and directly perform subsequent detection? to achieve lightweight,depthwise separable convolution and Re LU6 activation are used? in terms of data balance,Copy-Paste and Grid＿Mask are used to improve small target detection and generalization ability,and data augmentation methods such as stretching,masking,and flipping are used.The model’s good performance in lightness and robustness is verified on the BDD100 K real-time driving traffic dataset.(3)Constructing an online object tracker based on V-PGPR and the joint learning strategy.In order to accelerate the learning of the tracker in detection and appearance embedding,V-PGPR is used as the basic detector,and a part of the detection head is used for joint learning of detection and appearance features.In order to track faster,LAPJV is used instead of the Hungarian algorithm to solve the linear minimum cost allocation problem by searching for the shortest augmenting path,and to update the Kalman filter faster.The proposed tracker is tested on the KITTI traffic scene data,and the method performs well in terms of accuracy,and meets the speed requirements of real-time.Finally,visualization output and actual inference of the model on the NVIDIA workstation are performed. |
