Multi-Scene Traffic Light Detection and Fault Identification via Dual-Attention Image Fusion

Traffic light detection and fault identification using images from road traffic cameras are important for intelligent traffic management and urban safety monitoring. However, images collected in real traffic environments show clear differences in camera view, lighting conditions, weather, and background complexity. As a result, traffic lights vary greatly in scale, spatial location, and appearance, which reduces detection accuracy in complex scenes. To deal with this problem, this paper presents a multi-scene traffic light detection and fault identification framework based on dual-attention image fusion. Large-scale road camera data from the Chengdu Traffic Management Bureau are used, together with the Bosch Small Traffic Lights (BSTL) dataset. A traffic light and fault dataset is first built and expanded. Then, to improve detection under complex backgrounds and scale changes, a DCSPx module is designed to combine global features with local features in the backbone network. At the same time, a dual-attention mechanism is introduced. This mechanism includes position attention and channel attention, which helps improve effective feature learning. Based on these components, a Multi-Dual-Attention YOLOv5xp (DA-YOLOv5xp) algorithm is developed by combining detection results from images captured at different phases. In addition, a time-domain-based fault judgment method is proposed to detect abnormal traffic light states under real operating conditions. Experiments on real traffic camera data show that the method reaches an F1-score of 95.83% for traffic light detection and 96.97% for fault identification, and it performs better than several existing detection models.