A lightweight method for small scale traffic sign detection based on YOLOv4-Tiny

Automatic driving requires real-time consideration for traffic sign target detection algorithms while ensuring the accuracy. However, the current one-stage target detection algorithm mainly used for real-time detection is not focused on the characteristics of traffic signs, and the relevant research is insufficient. Aiming at this problem and ensure the accuracy of light-weight network in traffic sign detection task, an improved lightweight traffic sign recognition algorithm based on YOLOv4-Tiny was proposed, with improved backbone feature extraction and detection head using CBAM attention mechanism and depth-wise separable convolution, known as CDYOLO. Based on CDYOLO, we further proposed CDYOLO-SP, which can perform well in complex multi-category detection tasks. In terms of training methods, we adopt the transfer learning mode of "CCTSDB + TT100K" to improve performance. Compared with the original YOLOv4-Tiny, the improved algorithm has achieved better results. In the CCTSDB three-classification task, the mAP of CDYOLO improved by 6.52% and FPS maintained at about 82.5 FPS. The model size is only 4.1 MB. In the TT100K complex multi-classification task, the mAP of CDYOLO-SP improved by 48.59% and FPS maintained at about 60.2 FPS, and the model size is only 10.0 MB. Furthermore, the experiments show that compared with different CNN-based methods our methods outperforms them significantly. In summary, the improved model can meet the accuracy and real-time requirements of traffic sign detection and can be deployed on low-performance devices.