Research On The Methods Of Cyclist Detection In Traffic Scene

With the rapid development of urbanization,more and more people are convenient to travel by using cars as vehicles.However,as a vulnerable group of road traffic,a large number of pedestrians and cyclists die every year in traffic accidents,and the safety of the weak in the traffic is greatly threatened.Compared with the pedestrian detection,people have less research on the cyclist detection and the detection effect is worse.In order to reduce the incidence of traffic accidents and enhance the protection effect of the automatic driving system and the assisted driving system on the cyclist,this paper research on the method of cyclist detection in the traffic scene.And some detection algorithms are improved to enhance the detection effect.First of all,this paper research on the DPM algorithm which is the traditional detection methods.The DPM algorithm extracts the improved HOG features,which not only reduces the complexity of the algorithm,but also better describes the gradient direction characteristics of the target.The DPM algorithm uses multiple combined detection models to detect the target.The root filter is used to detect the general contour of the target.Then the part filters is used to detect the key parts.The final response is calculated by the combined response of the root filter and the part filters.And the response value is used to determine whether the detection area contains a target.The detection effect of the DPM algorithm is good,but the detection speed is too slow to be used for real-time detectionSecondly,this paper research on the cyclist detection algorithm based on deep learning,including YOLOv3 and SSD.YOLOv3 combines the advantages of YOLOv2,residual network and FPN network.YOLOv3 extracts target features by Darknet-53 and predicts the location and types of targets on the last three feature maps of different scales.SSD extracts target features by VGG16 and predicts the location and types of targets on the last six feature maps of different scales.The detection speed of these algorithms is fast.Finally,this paper makes two improvements to the SSD network model.One is we borrowed the idea of FPN algorithm,and using method of cross-layer connections in the different scale convolution feature layers in the SSD network,so that the network can not only learn high-level feature semantic information but also locate more accurately.And another is we use Focal Loss to calculate the confidence error to solve the problem of positive and negative sample imbalance during training and try to improve the detection effect of the SSD.