融合Transformer和改进PANet的YOLOv5s交通标志检测

Fusion Transformer and Improved PANet for YOLOv5s Traffic Sign Detection

针对交通标志检测速度慢和目标大小与类别极度不平衡等问题,提出一种融合Transformer和改进PANet网络的YOLOv5s交通标志检测算法。首先在不增加模型复杂度的前提下,将主干网络末端与Transformer融合以提高网络特征提取能力;其次由于所采用交通标志数据集的目标尺度太小,导致网络32倍大尺度检测层检测效果不佳,故不采用相关网络层,同时采用K-means算法得出适合的预测候选框;然后改进损失函数以解决正负样本极度不平衡问题。最后将所提出的改进算法在Jetson AGX Xavier平台上部署验证。实验结果表明,所提算法检测性能更佳,其准确率和召回率在原网络的基础上分别提高了2.2%和0.7%,模型参数量和计算复杂度分别减少了25.8%和10.1%。在Xavier上的检测速度达到76FPS,满足实时交通标志检测的要求且易于在实际场景部署。

Aiming at issues of low speed of traffic sign detection and extreme imbalance of target size and category,a YOLOv5s traffic sign detection algorithm of fused Transformer and improved PANet network is proposed.Firstly,without increasing model complexity,the end of backbone network is fused with Transformer to improve network feature extraction capability.Secondly,the small size of target scale in traffic sign datasets causes poor detection effects of the scale detection layers with a scale of 32 times larger,thus correlation network layers is not used,and K-means algorithm is adopted to yield appropriate prediction candidate frames.Then the loss functions is improved to address extreme imbalance problems between positive and negative samples.Finally,the proposed algorithm is applied on Jetson AGX Xavier platform for validation.Experimental results demonstrate that the proposed algorithm achieves better performance,the detection precision and recall rate improve 2.2%and 0.7%,respectively,and the number of model parameters and computational complexity reduce 25.8%and 10.1%respectively in comparison with the original network.The detection speed of 76 FPS on Xavier meets the requirements for real-time traffic sign detection,which is easily deployed in real scenarios.