Research On Pedestrian Detection Technology Of Road Traffic Environment Based On YOLOv3

Pedestrian detection in road traffic is an important task of environment perception for intelligent driving system,and also the basis of vehicle path planning,decision-making and control.The traditional pedestrian detection method and the deep learning based pedestrian detection method are compared by using the BDD100K(Berkeley Deep Drive)dataset.Considering accuracy and speed of target detection,we select YOLOv3 as the basic framework of pedestrian detection algorithm.A dataset named PD2018 about pedestrian for road traffic environment in Guangzhou is established through data collection,screening,preprocessing and labeling.In order to take the advantages of the YOLOv3 framework and utilize the computing resources,we optimize the network parameters and training parameters of YOLOv3.First,according to the prior knowledge of pedestrian dataset,the K-means++ clustering algorithm is used to obtain the optimal anchor box's width and height,and the optimal number of anchor boxes is determined by comparing the performance of models with different numbers of anchor box.Second,by comparing the effects on the model's floating-point computation and performance under different input image sizes,a model with the unique unequal input picture whose size of width and height are different is built.For the problem of low positioning accuracy of the model,the optimization is made to the training threshold of the prediction box in background and the weight of the coordinate prediction error in loss function.Since the size in the image of the pedestrian in the road traffic environment is small,a modified YOLOv3 is proposed.In the shallow layer of YOLOv3's basic network,a more detailed feature extraction layer is added.At the same time,the detection output of the network on the large-scale feature layer is increased,and the improved network model YOLO-M is obtained,and then verified by using the PD2018 dataset.The result shows that YOLO-M has improved detection accuracy under real-time detection.To improve the robustness and solve the missed detection problem of YOLO-M model,KCF filtering algorithm and Kalman filtering algorithm are introduced under different scenarios.Experiments show that the introduction of KCF filtering algorithm could reduce the missed detection rate of single pedestrian scenes and improve the detection speed,but the false detection rate is greatly affected by false detection of YOLO-M.The introduction of Kalman filtering algorithm would make the speed drop down,but could effectively reduce the missed detection rate of algorithms in multi-person scenarios.