YOLOv3 Vehicle Detection Method Based On DenseNet

As a basic component in driverless car system,vehicle detection algorithm has become a research hotspot.However,one of its main problems is that the limitations of imaging equipment and the impact of road environment often lead to imperfect vehicle detection results.Therefore,how to obtain a vehicle detection algorithm with high precision and high efficiency is the research goal of this thesis.The thesis proposes an improved YOLOv3 based vehicle detection algorithm.The main innovations are as follows:(1)A new Dense-YOLOv3 vehicle detection method is proposed with DenseNet as the backbone networks.Compared with ResNet as the backbone networks of the traditional YOLOv3,DenseNet uses concatenate and dense connection methods to retain the image feature information more effectively,thereby improving the detection accuracy;(2)A new loss function combining Focal Loss and GIoU Loss is proposed.Using Focal Loss as the confidence loss function can improve the detection performance on difficult samples;using GIoU Loss as the position information loss function can make the model more sensitive to the position information of bounding boxes,and the overlap conditions between predicted boxes and ground truth boxes,which can improve detection accuracy and convergence speed.Extensive experiments were carried on the BDD100K dataset.The experimental results show that the improved YOLOv3 algorithm based on DenseNet can achieve better detection results compared with the traditional YOLOv3 and other popular detection methods.