Real-time Detection Of Vehicle Targets Based On Improved YOLOv3 Algorithm

In recent years,automatic driver systems and advanced driver assistant systems have become the research hotspots of major automobile companies and Internet companies,and target detection is one of the core technologies.Although the existing target detection algorithms have made some progress in accuracy improvement,their large amount of parameters and high computational cost are not conducive to the application of low-storage devices.At the same time,urban traffic scenes are dense and complex,and vehicles are severely occluded,which often causes problems such as missed detection and false detection in target detection algorithms.This article first conducts research from the aspect of network lightweight.Through the optimization and improvement of YOLOv3,a lightweight vehicle detection algorithm is proposed,which makes the model smaller and the detection speed faster without reducing the accuracy.Then,from the perspective of reducing the missed detection rate and false detection rate,more targeted optimization strategies are designed for the loss function and non-maximum suppression algorithm to improve the applicability and robustness of the vehicle detection algorithm in dense scenes.The main contents of this paper are as follows:(1)In view of the uneven distribution of various types of vehicles in the BDD100 K data set,a random deletion operation is performed on cars and trucks with excessive data in the data set.At the same time,the data sets of bicycles and motorcycles are expanded using traditional data augmentation methods and data augmentation methods based on improved deep convolutional generative adversarial networks.Finally,a vehicle data set with a large amount of data,balanced categories and diverse scenes is established.(2)Aiming at the problem of the large amount of parameters and calculations in the real-time detection of vehicle targets by the YOLOv3 algorithm,a lightweight L-YOLOv3 algorithm is proposed.The residual unit of YOLOv3 is redesigned based on the depth separable convolution,SE module and Ghost module.The algorithm constructs a 16-layer backbone feature extraction network,retains the multi-scale prediction network of the feature pyramid structure,which together constitutes the lightweight L-YOLOv3 algorithm.The experimental results show that the L-YOLOv3 algorithm reduces the parameter amount by nearly 67% compared with the YOLOv3 algorithm,achieving the purpose of lightening the model.At the same time,the m AP value is increased by 3.18%,and the FPS is increased by 8 frames per second.It is better than the YOLOv3 algorithm and improves the overall performance of the vehicle target detection algorithm.(3)In response to the problem that the traditional NMS algorithm can only retain one candidate frame when multiple targets overlap,the SD-NMS post-processing method is proposed.This method effectively combines the idea of multi-target set prediction with DIo U,which simplifies the feature learning process.The experimental results show that the missed detections and false detections in the dense vehicle scene are effectively reduced.At the same time,in order to improve the convergence speed and accuracy of the target frame regression,the CIoU Loss is introduced in the loss function optimization.It is concluded from the experimental results of regression loss that the loss curve of the L-YOLOv3 algorithm after using the CIoU Loss is more stable during the target frame regression process.