MA-YOLO: a Lightweight Vehicle Detection Framework Based on YOLO

The current vehicle detection models are inadequate for deployment on edge computing devices due to limited computing capabilities. Therefore, this paper proposes a novel lightweight vehicle detection method called MA-YOLO (MobileNet Attention YOLO). In order to reduce the number of parameters and computational requirements, the depthwise separable convolution and inverse residual structure is used to instead of conventional convolution. Additionally, the SELayer and CALayer module of attention mechanism is embedded into Backbone module, to improve the correlation of inter-channel and inter-space features, respectively. The experimental results demonstrate that the MA-YOLO has achieved a mean average precision (mAP) of 90.7% for vehicle detection under normal weather conditions. In comparison to YOLOv8s, MA-YOLO has compressed the number of parameters by 40.5% and computation by 52.4%, while experiencing only a decrease of 0.1% in mAP.