SAM-YOLO: An Improved Small Object Detection Model for Vehicle Detection

Vehicle detection using computer vision plays a crucial role in accurately recognizing and responding to various road conditions, targets, and signals, particularly within autonomous driving technology. However, traditional vehicle detection algorithms suffer from slow detection speed, low accuracy, and poor robustness. To address these challenges, this paper proposes the simple attention mechanism-you only look once (SAM-YOLO) algorithm. SAM-YOLO incorporates the simple attention mechanism into the YOLOv7 network, allowing for the capture of more detailed information without introducing additional parameters. In this study, we experimentally redesigned the backbone network of SAM-YOLO by replacing the redundant part of the network layer with the C3 module, resulting in improved model performance while maintaining accuracy. The experimental results show that the SAM-YOLO algorithm performs excellently in several evaluation metrics under conventional conditions, especially outperforming other algorithms in accuracy and mean average precision values. In tests on the ExLight dataset facing extreme lighting conditions, SAM-YOLO similarly demonstrated optimal detection capabilities, especially in terms of robustness when dealing with complex lighting variations. These findings emphasize the potential of the SAM-YOLO algorithm for real-time and accurate target detection tasks, especially in environments with highly variable lighting conditions.