"基于多尺度YOLOv5的高效实时交通灯检测算法优化研究"

Traffic light detection is a crucial task in the field of autonomous driving, as it directly impacts the safety of vehicles on the road. The difficulty in detecting traffic lights due to their small scale and complex environmental conditions has prompted the development of a multi-scale YOLOv5 traffic light detection algorithm. This approach involves utilizing a combination of data augmentation techniques to enhance the complexity of the model input, as well as training the model at multiple scales instead of a fixed scale to improve its learning capabilities. Additionally, a multi-scale feature fusion network is constructed to merge information from downsampled scales of 4x, 8x, 16x, and 32x, creating a multi-scale detection layer. To enhance the feature fusion capability, long-range skip connections are introduced to transfer information between different levels, significantly improving the model's ability to detect small objects. Experimental results demonstrate that the improved YOLOv5 algorithm achieves a detection speed of up to 9.5ms and an mAP of 99.8% on the collected dataset, representing a 17% improvement over the original YOLOv5 model. Furthermore, on the Bosch dataset, the mAP increases by 6.5%, enabling real-time and high-precision traffic light detection.