基于YOLOv7的交通目标检测算法研究

Research on Traffic Object Detection Algorithm Based on YOLOv7

摘要:

针对交通场景中由光照、遮挡、目标小以及背景复杂等因素导致目标检测精度低，易出现漏检和误检问题的情况，提出了一种基于YOLOv7的交通目标检测算法；该算法在主干网络中融入多头注意力机制，以增强网络特征学习能力，从而更好地捕获数据和特征内部的相关性；在YOLOv7颈部网络引入协调注意力模块(CA)，将位置信息嵌入到注意力机制中，忽略无关信息的干扰，以增强网络的特征提取能力；增加一个多尺度检测网络，以增强模型对不同尺度目标的检测能力；将CIoU损失函数更改为SIoU函数，以减少模型收敛不稳定问题，提高模型的鲁棒性；实验结果表明，改进的算法在BDD100K公开数据集上的检测精度和速度分别达到了59.8% mAP和96.2 FPS，相比原算法检测精度提高了2.5%；这表明改进的算法在满足实时性要求的同时，具备良好的检测精度，适用于复杂情况下的交通目标检测任务。

Abstract:

Aiming at the situation that the target detection accuracy is low due to the factors of lighting, occlusion, small target and complex background in complex traffic scenes, and is prone to missed and false detection, a traffic target detection algorithm based on YOLOv7 is proposed. To better capture the correlation within data and features, the algorithm incorporates a multi-head attention mechanism into the backbone network to enhance the network feature learning ability. The Coordinated attention module (CA) is introduced into the YOLOv7 neck network and the position information is embedded into the attention mechanism, which can ignore the interference of irrelevant information and enhance the feature extraction ability of the network. A multi-scale detection network is added to enhance the detection capability of the model for different scale targets. Changing the CIoU loss function to SIoU function to reduce the problem of model convergence instability and improve the robustness of the model. Moreover, the results show that the detection accuracy and speed of the improved algorithm on the BDD100K public dataset reach 59.8% mAP and 96.2 FPS, respectively, representing an increase of 2.5% compared to that of the original algorithm. It shows that the improved algorithm has good detection accuracy while meeting the real-time requirements, which is suitable for traffic target detection tasks in complex situations.