TSM-YOLO: An Optimized YOLOv11 Model for Traffic State Surveillance

With the increase in traffic flow and the growing complexity of traffic scenes, constructing a precise and efficient traffic state monitor has become a core topic in intelligent transportation. Traffic state monitor involves multi-target recognition, positioning, and status monitoring, which plays a crucial role in aspects such as traffic management. The YOLOv11 model has excellent real-time detection capabilities, but it has limitations when used for traffic state monitor. This paper proposes the TSM-YOLO target detection model, which innovatively optimizes the key aspects of the original YOLOv11 to improve its detection performance in traffic scenes: 1.Regarding the MPDIoU function, The original IoU metric lacks sufficient precision and thus cannot measure the matching degree between the predicted box and the real box, especially in complex situations where traffic targets are occluded. By analyzing the spatial distribution and mutual relationships of traffic targets, this function is improved and penalty mechanism are optimized, the accuracy of target positioning is enhanced, and the detection error is reduced. 2. Given the importance of feature extraction, the SENetV2 algorithm is introduced. Traffic targets are diverse, and traditional feature extraction modules have difficulty adapting to capture key features. However, SENetV2 can adaptively recalibrate feature channels to highlight important features. For example, when distinguishing vehicles, it can strengthen the key feature channels to make classification and positioning more accurate. To verify its effectiveness, a large-scale dataset contain multiple actual traffic scenes, and comparative experiments were carried out after annotation and pre-processing. The results show that compared with the original algorithms model, TSM-YOLO algorithm has a significantly improved average accuracy rate. It can detect key targets more accurately to provide strong support for traffic management and research value.