Research On Vehicle And Pedestrian Detection Algorithm Based On Deep Learning

The development of auto industry and information technique has made it possible for the implementation of unmanned driving.At present,many well-known car companies in the world have carried out the strategic layout on self-driving.Unmanned driving has been carried out experimentally recently.The perception system is a crucial component of the self-driving system,which provides the traffic environmental information for following decision-making system.Image target detection is a crucial technologies in perception system.Object detection technology based on deep learning has been a research focus recently,new algorithms with outstanding performance are emerging.They gradually replace the traditional detection algorithms and are widely applied in production and life.According to specific steps,object detection algorithms of deep learning can be categorized into two-stage algorithms and onestage algorithms.The two-stage algorithms represented by R-CNN firstly extracts object proposals from the image and then perform location and classification regression to get the detection targets.They have high detection accuracy,but the poor speed makes them unsuitable for real-time detection,in unmanned driving.The one-stage algorithms represented by YOLO directly use the end-to-end method to get the detection results.The one-stage algorithms are better than the two-stage in detection speed and are widely used in practical applications.This thesis focuses on the vehicle and pedestrian detection based on YOLOv3.To deal with the high model complexity and computational cost of YOLOv3 and the difficulty of detection increasing caused by complex traffic environment,three improvement strategies are proposed to optimize the vehicle and pedestrian detection based on yolov3.(1)Aiming at the abundant parameters,high computational cost and large runtime memory of the CNN,a lightweight vehicle and pedestrian detection algorithm based on YOLOv3 is proposed.Through adding L1 regularization of scaling factor in BN layer to the loss function of YOLOv3 sparsity training is carried out.Then the channels and residuals module with too small factor are pruned to get the compact model,the parameters and computational cost are decreased and detecting speed is increased.(2)The problem of missing detection is serious due to too much occlusion in traffic environment.In the post-processing stage,the traditional NMS algorithm is replaced by the Merge Soft-NMS algorithm proposed on the basis of Soft-NMS in this thesis.The score of the non optimal prediction boxes is decayed rather than suppressing it altogether,and the prediction boxes with a high overlap of the same target are merged according to the score weight.The Merge SoftNMS reduces the problem of missing detection and improves the robustness of localization.It does not need to be retrained and implement easily.(3)The anchor boxes in YOLOv3 is extracted from multi categories dataset by K-means clustering algorithm,which are unsuitable for vehicle and pedestrian.To avoid the increasing of regression difficulty caused by inappropriate anchor boxes and the unstable K-means algorithm in random starting points,this thesis uses K-means++ algorithm to extract the appropriate anchor boxes from vehicle and pedestrian dataset.The proposed algorithm is tested on the detection dateset of unmanned driving to examine the effectiveness.The results indicate that the proposed algorithm has significant superiority in speed,fewer parameters and computational cost with the accuracy ensured.