Research on CEL-YOLO Algorithm for Lightweight Detection of Traffic Signs

Nowadays, in the era of rapid development in the field of intelligent transportation and large information reserves, in order to better adapt to embedded devices and improve the real-time and robustness of intelligent vehicle perception, this paper proposes a traffic sign small target detection algorithm based on YOLOv8 model. First, on the basis of C2f-STAR module, StarNet convolution StarsBlock is added to build C2f-Starsblock, which replaces C2f module in Backbone part of YOLOv8 network to improve the feature representation capability and detection performance of the model. Secondly, based on the BottleNeck of the residual module Faster_Block in FasterNet, the C2f module in YOLOv8 network is replaced, and EMA attention mechanism is added to the C2F-FASTER module to construct the C2f-Faster-EMA module. Improve the ability of C2f module to capture multi-scale feature information; Thirdly, the SPPF module is combined with the large separable Kernel Attention (LSKA) module to construct the SPPF-LSKA module to enhance the feature extraction capability of the model. Finally, a Light Weight Shared Convolutional Detection (LSCD) is added. It can be seen in the CCTSDB2021 traffic sign dataset, finally the improvement of the traffic sign this kind of small target detection accuracy and robustness of the model. To verify the effectiveness of CEL-YOLO, mAP-50 achieves 97.1% in traffic sign detection tasks. When the total number of parameters and calculation amount are reduced by 54.5% and 44.5% respectively, the accuracy remains the same as the original model. The verification results show that compared with the benchmark model, the model is significantly lighter in volume and computation, and is more suitable for small target detection.