针对复杂交通场景下，特别是拥堵道路中，经常出现的交通目标密集、互相遮挡，小尺度目标检测精度低的问题，提出了一种面向复杂交通场景的目标检测模型YOLO-T(You Only Look Once-Transformer)。首先提出CTNet主干网络，相较于CSPDarknet53，该主干拥有更深的网络结构和多尺度特征提取模块，不仅能够更好地学习密集目标的多级特征，还可以提高模型对复杂交通场景的应对能力，进而引导模型更加关注小目标的特征信息，提升小目标的检测性能;其次引入Vit-Block，采用卷积和Transformer并行的方式融合更多的特征，兼顾局部和上下文信息的关联性，从而提升检测精度;最后在颈部网络Neck后增加Reasonable模块，引入注意力机制，进一步提高目标检测算法对复杂场景和遮挡目标的鲁棒性。实验结果表明，相比基准算法，YOLO-T 在 KITTI 数据集和 BDD100K 数据集的检测精度分别提高了 1。92%和12。78%，能有效提升复杂交通场景下的检测性能，更好地辅助驾驶员对其他车辆行驶行为的判断，减少交通事故的发生。
Object detection model YOLO-T for complex traffic scenarios
To address the challenges posed by complex traffic scenarios,particularly congested roads where traffic objects are densely packed and often occlude each other and small-scale objects are detected inaccurately,a new object detection model called YOLO-T(You Only Look Once-Transformer)is proposed.First,the CTNet backbone network is introduced,which has a deeper network structure and multiscale feature extraction module compared with CSPDarknet53.Not only can it better learn the multilevel features of dense objects but can also improve the model's ability to handle complex traffic scenarios.Moreover,it directs the model's focus toward the feature information of small objects,thereby improving the detection performance for small-scale objects.Second,Vit-Block is incorporated,which integrates more features by parallelly combining convolution and Transformer.This approach balances the relevance of local and contextual information,thereby enhancing detection accuracy.Finally,the Reasonable module is added after the Neck network,introducing attention mechanisms to further improve the robustness of the object detection algorithm against complex scenarios and occluded objects.Experimental results indicate that compared with baseline algorithms,YOLO-T achieves a 1.92%and 12.78%increase in detection accuracy on the KITTI and BDD100K datasets,respectively.This enhancement effectively boosts detection performance in complex traffic scenarios and can assist drivers to better predict the behaviors of other vehicles,thus reducing the occurrence of traffic accidents.
国家科技期刊平台
NSTL
万方数据
