Research on pedestrian detection algorithms combined with lightweight networks

Given the poor real-time performance and the lack of precision in the current YOLOv4 algorithm for pedestrian target detection, an improved algorithm based on YOLOv4 is proposed. Firstly, replace the backbone feature extraction network of the original YOLOv4 algorithm with the lightweight network GhostNet, while depthwise separable convolution is used to replace the normal convolution in the enhanced feature fusion network, which simplifies the model and significantly reduces the computational effort, thus speeding up detection and improving real-time performance. Secondly, CA(Coordinate Attention)focus mechanism is incorporated into the enhanced feature extraction network of the YOLOv4 algorithm to enhance the feature representation capability of the network model, the Focal Loss function was also introduced for resolving either positive or negative sample imbalance to improve detection accuracy for the algorithm. The experimental findings have shown that the size of the improved algorithm model has been reduced from 245.53MB to 43.64MB and the memory footprint has been decreased by 82%; the speed of detection reached 54.24 FPS, an improvement of 18.87 FPS compared to the original algorithm; the average precision of detection reached 93.28%, an improvement of 2.72% over the original algorithm. The improved algorithm achieved faster detection speeds, better real-time performance as well as better detection results in pedestrian target detection tasks.