Research On Target Detection Of Driverless Scene Based On YOLO Algorithm

With the deepening of artificial neural network research and the rise of convolutional neural network research,computer vision has gradually entered the public vision.With the expansion of computer vision research,target detection,as an important research direction of computer vision,has gradually played an important role in production and life.With the development of 5G technology,automatic driving has become the object of the development of governments and technology companies all over the world,and target detection in unmanned automatic driving system has become the focus of researchers.But under the condition of driving scene,speed and accuracy are the most important standards in unmanned target detection,so the future optimization direction of unmanned driving must be detection speed,detection accuracy and detection missing rate.There are two main directions for target detection in driving scenarios: target type detection and target position detection.Due to the excellent image processing capability of the convolutional neural network(CONVOLUtional neural network),combined with the comparison of several excellent image recognition algorithms at present,this paper adopts this algorithm to optimize the yolov3-tiny and Yolov4-Tiny algorithms based on which the speed,accuracy and missed detection rate target detection driving scenes are optimized.Specific work contents include:(1)This paper presents an optimization method for YOLOv3-tiny algorithm:increase the network depth,increase the space pyramid pooling structure,and increase the yolo detection layer.First,YOLOv3-tiny algorithm is YOLOv3 lite version,YOLOv3 backbone network is Darknet-53,and YOLOv3-tiny backbone network is Darknet-19 structure,compare the speed above to get a lot of ascension,but it is on the basis of sacrificing the accuracy and the miss rate,this paper SPPNet model space in the pyramid model,join in YOLOv3-tiny algorithm backbone network is similar to the pyramids,the space SPPNet model of pool model,It solves the problem of repetitive feature extraction in convolutional neural network,and the increase of network depth is conducive to the sampling and calculation of eigenvalues,greatly improves the generation speed of candidate box,and saves the calculation cost without sacrificing the calculation speed.Second,YOLOv3-tiny has only two YOLO layers.In order to increase the accuracy and diversity of detection,a layer of YOLOv3-tiny has only two layers of YOLO layers.(2)Combined with the improved research on YOLOv3-tiny algorithm,the optimization of the new algorithm,YOLOv4-Tiny algorithm,is achieved by adding YOLO layer and adopting more efficient Mish activation function.The yolov4-tiny backbone network is CSPDarknet-19 structure,and the YOLOv3-tiny algorithm has only two YOLO layers.This paper also adds YOLO layer on this basis.In a convolution neural network,a stable activation function with output results is more conducive to the downward transmission of eigenvalues and the extraction and computational sampling of eigenvalues at each layer.As Leaky-relu activation function in the previous YOLO-tiny algorithm is abandoned in this paper,a more stable Mish activation function is adopted to improve the overall accuracy and leakage rate.Because the safety of driving scene mainly based on the safety of the people as a consideration,based on the principle of people-oriented,must strengthen the elements for human detection,in actual driving scene both pedestrians and cyclists are based on human factors is given priority to,in this case,this paper introduced the special change for data collection,the human factor related species are classified as pedestrians,classify and in image annotation,reduce kinds but can increase the data volume,better improve accuracy.