SMA-YOLO: A Novel Approach to Real-Time Vehicle Detection on Edge Devices

Vehicle detection plays a pivotal role in traffic management as a key technology for intelligent traffic management and driverless driving. However, current deep learning-based vehicle detection models face several challenges in practical applications. These include slow detection speeds, large computational and parametric quantities, high leakage and misdetection rates in target-intensive environments, and difficulties in deploying them on edge devices with limited computing power and memory. To address these issues, this paper proposes an improved vehicle detection method called SMA-YOLO, based on the YOLOv7 model. Firstly, MobileNetV3 is adopted as the new backbone network to lighten the model. Secondly, the SimAM attention mechanism is incorporated to suppress background interference and enhance small-target detection capability. Additionally, the ACON activation function is substituted for the original SiLU activation function in the YOLOv7 model to improve detection accuracy. Lastly, SIoU is used to replace CIoU to optimize the loss of function and accelerate model convergence. Experiments on the UA-DETRAC dataset demonstrate that the proposed SMA-YOLO model achieves a lightweight effect, significantly reducing model size, computational requirements, and the number of parameters. It not only greatly improves detection speed but also maintains higher detection accuracy. This provides a feasible solution for deploying a vehicle detection model on embedded devices for real-time detection.