Road Vehicle Detection Based on Deep Learning

To address the challenges of low accuracy in vehicle detection for small targets, occluded targets, and scenes with limited lighting conditions, this paper proposes a vehicle detection method based on YOLOv7. Firstly, a pyramid pooling structure incorporating depth-wise separable and dilated convolution modules is constructed to allow the model to integrate contextual information while enlarging the receptive field and reducing the parameter count. Next, parallel channel-spatial attention modules are designed to eliminate interference caused by serial connections. Experimental results demonstrate that the proposed detection model achieves an average precision (mAP) of 98.9% and 83.4% for mAP50 and mAP50:95, respectively. Compared to YOLOv7, this represents an improvement of 2 and 5.8 percentage points, while achieving a processing speed of 39 frames per second.