Research On Vehicle Detection And Tracking Algorithm Based On Instance Segmentation

In recent years,autonomous driving has become a central issue in the field of artificial intelligence.The model of vehicle detection and tracking algorithm with superior performance is designed and applied to autonomous driving system,traffic monitoring system and intelligent parking system,which has become the main research direction of computer vision.Based on deep learning,this paper studies vehicle detection and tracking algorithm,and uses deep network to extract vehicle features,which not only meets the requirements of real-time detection,but also achieves high detection accuracy.The main innovations of this work include the following:In terms of vehicle detection,in this work,we use state-of-the-art one-stage approach YOLACT,which is a fast and real-time instance segmentation model algorithm.For the backbone detector,the Res Net-50 network is adopted to extract more abundant semantic feature information of the target object.In the aspect of target positioning,this paper proposes a more precise P-CIo U loss function to predict the contour of object as far as possible.In the P-CIo U loss function,distance constraint parameter is introduced to control the scale regression parameter,which can affect the value of the entire regression loss,so that the prediction box makes more accurate regression target box.The experimental results demonstrate that the improved P-CIo U loss has an improvement of about 0.6% in AP compared to the CIo U loss.Moreover,in the final candidate box selection,this paper proposes a Skip-calculation NMS mechanism,which combines Gaussian penalty function and distance constraint parameter introduced in P-CIo U loss to effectively solve the problem of target with occlusion being suppressed.Meanwhile,when multiple objects are close and occluded,the improved Skip-calculation NMS algorithm can appropriately increase the penalty score of prediction boxes adjacent to target box with the highest score.This algorithm is beneficial to improve confidence score of prediction boxes and guarantees the detection rate of targets to a certain extent.Compared with Fast NMS algorithm,its average precision is improved by about 1.7%.In terms of vehicle tracking,this paper discusses three Siamese network object tracking algorithms: Siam FC,Siam RPN as well as Siam Mask and reimplements Siam Mask basic network and mask refinement network.In order to verify the performance of Siam Mask,this paper chooses two widely used datasets: VOT-2016 and VOT-2018 to test the video frame sequence of the vehicle.The experimental results show that the tracking speed is up to 75 FPS,which meets the real-time requirement.At the same time,this paper discusses four different vehicle video scenarios from VOT-2018 dataset for qualitative analysis.This paper further analyzes the influence of object occlusion and background blur on the stability and accuracy of the tracker,and concludes the defects of existing framework.