| The development of the nation’s economy and the advancement of urbanization have caused a sharp increase in the number of cars in many cities.In the meantime,problems such as traffic jams and frequent car accidents have also emerged.The traffic carrying capacity and management capabilities of major cities have reached a bottleneck,while traditional solutions’ hard to solve the problems.Intelligent traffic management and autonomous driving technologies provide innovative solutions to urban traffic problems.Computer vision technology is a key cornerstone of intelligent traffic management and autonomous driving technology,and object detection technology is an important part of it.As object detection technology based on convolutional neural networks is at the top of the leading board,it is of great significance to design an object detection algorithm model that can perform superiorly on the traffic object data set.This paper will focus on mixed domain attention and multiscale features.The main contents are as following:(1)Firstly we studied fusion methods of CBAM.We propose using a nonlinear mapping to adjust the signal in the attention map generated by the CBAM attention module to solve the problem that the parallel fusion strategy has a depressing effect on the useful signal.With this nonlinear mapping,the top-1 error for the Image Net-250 validation set of Res Net50 which adopting the arithmetic average fusion strategy and the geometric average fusion strategy is reduced by 0.31% and 0.4%,respectively.AS for GSo P,we propose an improved module named GSo P-dw using depthwise separable convolution to fully integrate the context information of its channel dimension and space dimension separately.This fusion method makes the GSo P-Net’s top-1 error on Image Net-1K validation set dropped by 0.66%.In addition,we plug the GSo P-dw attention module into the object detection algorithm model YOLOv5 to enhance the ability of extracting useful features and improve the model’s accuracy on the traffic object detection data set BDD100 K.Furthermore,according to the results of experiments,the best attention module embedding scheme is also confirmed.(2)Then,in order to strengthen the ability of utilizing multi-scale features,we propose a new module SPP-Shuffle using channel shuffling operation and group convolution to replace the original SPP module which contains inefficient crosschannel convolution operation,fusing multi-scale receptive field information in the model.Experiments with YOLOv5 model show that using SPP-Shuffle to replace the original SPP can reduce the amount of calculation and the amount of parameters and improve the accuracy of the model.The basic component module CSP-block in the backbone network is improved to the Res2 CSP module to improve its ability to utilize multi-scale features.Through the above method,we increased the m AP@0.5 of the model on the BDD100 K validation set by 1.48,while maintaining the speed and hardware-friendliness of the model. |
