Driverless Vehicle Detection Based On Deep Learning

The increasing number of cars brings convenience to people,but also brings many challenges,such as: traffic congestion,environmental pollution,the increasing incidence of car accidents and so on.Driverless technology is the key to solve these problems,so it has gradually become the focus of attention.Vehicle detection,as a basic but very important part of driverless technology,has a high research value.At present,the vehicle detection method mostly adopts the method of multi-sensor fusion,but some sensors,such as lidar,are expensive,which limits the popularity of driverless technology,while the camera cost is relatively low,and it is easier to obtain and interpret visual clues,so this paper mainly studies the vehicle detection solution using camera.First,the vehicle detection algorithm-pixels for predict network.Considering the difference between segmentation task and detection task,this paper adds Salieny Map stage to detect the attention area pixel by pixel.Further analysis of the FCOS algorithm Center-ness suppression of low-quality bounding box deficiencies,inspired by the human posture key point detection algorithm PAF,the pixel point as the detected human body key point,the introduction of Center-surround as a new constraint method.Compared with FCOS,Salience Map can greatly reduce the amount of computation;Center-surround can reduce the number of pixels far away from the center of the object,and restrict the detection area near the center of the object,which has a better effect than center ness.The performance of KITTI and COCO data sets surpasses Faster R-CNN,SSD,YOLO and other algorithms.Secondly,vehicle tracking algorithm-violence update algorithm and Master-Worker tracking algorithm.In view of the shortcomings of the existing tracking algorithm,a method combining detection and tracking is proposed.Violence update algorithm is based on feature matching between detection region and tracking region,and information updating is based on feature results.However,the continuity of information in the matching time interval is not considered,so a master worker algorithm based on Kalman filter is proposed,whose allocation and detection mechanism ensures the continuity of information in the time sequence.Experimental results show that the violence update algorithm can update information correctly,but the conditions are limited.The master worker algorithm ensures the continuity of information through dynamic allocation,and its performance is better than the violent update algorithm.The performance of the two algorithms is better than that of Boosting,KCF and MOSSE tracker.