YOLOv7-VD: An Algorithm for Vehicle Detection in Complex Environments

Addressing the problems that existing vehicle detection algorithms do not have high detection accuracy in complex scenes, as well as the problem that the parameters of the deep learning model are large and occupy too much computer video memory, an improved algorithm, YOLOv7-VD, based on YOLOv7, is proposed. First, to simplify the operation of the network, the ghost convolution module will be introduced on the feature extraction network. Second, the spatial pyramid pooling cross stage partial channel-channel shuffle (SPPCSPCS) module is proposed. To improve the fusion of feature maps with different depths, this module adds channel shuffling to the spatial pyramid pooling structure. Once again, the network utilizes a 9x9 convolution kernel to replace certain 3x3 convolution kernels. This is done to enhance the network's active sensory patch and shape bias, and a 1x1 convolution is used to adjust the input feature layer to reduce model size and delay. Finally, the article introduces the Convolutional Block Attention Module (CBAM) attention mechanism that enhances the feature representation of the target being detected while removing redundant convolutional kernels and reducing parameters. The experimental tests demonstrate that the enhanced YOLOv7 improves mAP@0.5 by 2.01% over the baseline algorithm on the BDD 10K dataset, improves accuracy by 5.18%, and reduces the total number of parameters by 15.07%.