Research On Preceding Vehicle Detection Optimization Algorithm Based On Faster R-CNN

In recent years,the deep learning algorithm such as Faster R-CNN(Faster Region with Convolutional Neural Network)has been demonstrated successful for the detection of preceding vehicles on the road(referred to as "preceding vehicles"),but also faces some challenges caused by the uncertainty of vehicle movement、occlusion and other factors.In this paper,based on the Faster R-CNN algorithm,we will explore the problems of detection accuracy of preceding vehicles、the occluded vehicles detection and the detection speed of preceding vehicles,and explore the optimization of corresponding Faster R-CNN algorithm.Firstly,in response to the situation that the occurrence frequency of preceding vehicles with different sizes in image is constantly changing due to the speed change of autonomous vehicle(camera installation vehicle),the vehicle speed is divided into three stages:0-20km/h、20-60km/h、60-120km/h,and then at different speed stages,K-means clustering method is used to analyze the occurrence frequency of different sizes preceding vehicles,this paper finds that with the autonomous vehicle speed increasing,the occurrence frequency of small vehicles increasing and other types vehicles decreasing,thus this paper proposes the speed classification random anchor method,according to this rule,the size of anchor is re-optimized to improve the detection accuracy of Faster R-CNN algorithm for preceding vehicles.Secondly,for the detection of occluded vehicles,the NMS(Non Maximum Suppression)algorithm in Faster R-CNN is replaced by Soft NMS algorithm and the vehicle detection is carried out.According to the detection results,this paper finds that the penalty coefficient in Soft-NMS can be further optimized,thus the Q-square penalty coefficient method is proposed to multiply the penalty coefficient many times,the detection accuracy of Faster R-CNN for occluded vehicles is improved;at the same time,two new penalty coefficients of inverse proportion and index are applied in Soft NMS,and the influence of the new penalty coefficient on the detection accuracy of occluded vehicles is explored.Finally,for the detection speed of Faster R-CNN algorithm for preceding vehicles,this paper finds that the post processing method of RPN(Region Proposal Network)can be further optimized.First of all,the number of proposal areas and the number of detection boxes which selected according to score are adjusted,and then the new selection conditions of the detection box are added,by reducing the non vehicle shape detection boxes,the calculation amount is reduced.The experimental results show that the algorithm can greatly improve the detection speed of the preceding vehicles under the appropriate number and parameters selection.