Traffic Sign Detection Algorithm Based on CA-YOLOV4-tiny
Jinghan Xu, Pengzhi Chu, Xiongziyan Xiao, Hongcheng Huang · 2023
Autonomous driving technology is gaining popularity, but most deep learning algorithms for target detection of traffic signs suffer from a large number of parameters and a long inference time. This paper proposes a CA-YOLOV4-tiny traffic sign detection algorithm based on the YOLOV4-tiny network, feature extraction is performed by incorporating the attention mechanism, mosaic data is used to improve the generalization ability of the model, and label smoothing is used to prevent overfitting. The mAP on the TT100K traffic sign dataset is 81.71%, which is 6.77% higher than that of YOLOV4-tiny. The model is deployed on the NVIDIA Jetson Xavier NX artificial intelligence computing platform with a frame rate of 28 FPS, and the frame rate is 60 FPS after the model is quantif ied and accelerated using TensorRT. The experimental results show that CA-YOLOV 4-tiny can significantly improve the model detection accuracy without reducing the reasoning speed, and the reasoning speed after quantization has also been significantly improved, resolving the vehicle-side target detection network's problem of excessive computing power and slow reasoning speed.