Improved Traffic Sign Detection Model Based on YOLOv7-Tiny

Traffic sign detection is a critical task in the autonomous driving. Ordinary networks cannot obtain satisfactory results in traffic sign detection because the size distribution of traffic signs are extremely unbalanced. To overcome this challenge, this paper proposed an improved YOLOv7-Tiny object detection model. Firstly, a path connection strategy was proposed to enhance small-scale feature representation. Compared to the original FPN connection strategy, it adds a path that leads out of the backbone and connects into the Feature Pyramid Network(FPN). Secondly, we proposed a new down-sampling module--Slice-Sample. By slicing, the size of the feature map is reduced and subsequently, the weights of the sliced feature map channels are assigned using the channel attention mechanism. It can reduce the loss of feature information. Additionally, a module for detecting attention was proposed to address the aliasing effect found in the fusion of different scales. This channel attention mechanism not only focuses on the correlation of neighboring channels, but also employs two branches to increase the model's ability to extract information from the feature map. Experiments on the German Traffic Sign Detection Benchmark (GTSDB) showed that the improved model can achieve more remarkable performance than yolov7-tiny. Our method achieved 93.47% mean average precision (mAP) surpassing the yolov7-tiny's 7.48%, and the frames per second (FPS) value is maintained at 67.5. Besides, our method is superior to other lightweight models on the GTSDB. To demonstrate the generalizability of our approach, we tested it on the Tsinghua-Tencent 100K dataset (TT100K) without tuning and obtained 66.29% mAp surpassing the yolov7-tiny's 7.59%. In addition, the number of parameters of improved YOLOv7-Tiny is about 23.29 M.