Research On Vehicle Detection Algorithm Based On Deep Learning

With the rise of 5G technology,unmanned vehicles have begun to experimentally land,and autonomous driving has become a hot topic in the technical field.Imagebased vehicle detection is one of the basic technologies in autonomous driving.The use of on-board cameras to capture images of vehicles in front and detect vehicles in images is particularly important for autonomous driving applications.However,in complex scenes,because the vehicles in the images often appear occluded cases and the vehicles in the distance are small,etc.,this still poses a great challenge to the accuracy and robustness of the detection algorithm.With the continuous development of deep learning during near years,vehicle detection algorithms based on deep learning have also become a popular research direction.In the current research at home and abroad,vehicle detection algorithms based on deep learning are mainly divided into two categories: two-stage detection algorithms and one-stage detection algorithms.The two-stage detection algorithm represented by Faster R-CNN divides the entire detection process into two stages: region proposal,classification and regression.The one-stage detection algorithm represented by YOLO and SSD uses an end-to-end method to directly obtain the prediction boxes.Because the one-stage detection algorithm has the advantage of fast speed,it can take into account the real-time performance with higher detection accuracy,so it is more realistic and welcomed by scholars.This paper focuses on vehicle detection based on one-stage detection algorithms.After reviewing the research background,research status and basic theory,two improved vehicle detection algorithms are proposed.Aiming at the problems of insufficient feature extraction capability of the backbone network and insufficient detailed screening of prediction boxes,an improved vehicle detection algorithm based on YOLOv3 was proposed.This algorithm reinforces the feature extraction capability of the backbone network by introducing SE blocks in the backbone network Darknet-53 and using the attention mechanism to strengthen the information between the channels of the feature maps.In the prediction stage,the Gaussian-weighted Soft-NMS algorithm is used to replace the conventional NMS method,thereby effectively improving the screening process of the prediction boxes,and this improvement does not require retraining the model.Experiments show that this algorithm is better than the existing detection algorithms based on YOLO series.Aiming at the problems of low matching degree between the preset prior boxes and the vehicle dataset,insufficient feature extraction capability of the prediction network,and insufficient penalty of the loss function,a vehicle detection algorithm based on improved RFBNet was proposed.First,the K-means algorithm is used to select the prior boxes suitable for the vehicle dataset,so as to avoid the problem of inappropriate prior boxes caused by manual selection.Secondly,the RFB module in the prediction network is improved.By adding additional convolution branches to increase the field of view,the network’s feature extraction capability is improved.Finally,the box regression loss function is improved,and the CIoU loss is used to improve the penalty effect of the regression bias.Experiments show that the algorithm is superior to the existing SSD-based detection algorithms.