Attention module performance.

Target detection is an essential direction in artificial intelligence development, and it is a crucial step in realizing environmental awareness for intelligent vehicles and advanced driver assistance systems. However, the current target detection algorithms applied to complex real-life scenarios still have a lot of intractable problems, such as the detection of different road scenarios, not having a good real-time detection capability, and so on. Therefore, there is a need to balance the efficiency and effectiveness of the target detector. YOLOv7, as a single-stage target detection algorithm, combines a number of advanced modules and methods for the purpose of achieving more precise and faster target detection. This paper, YOLOv7 is used as a baseline, combined with deformable convolution, to realize adequate detection in complex scenes and balanced computational efficiency and accuracy by adding an attention mechanism module. In addition, combining the lightweight network module speeds up the model’s computational speed while improving the detector’s feature expression ability, thus accomplishing the task of real-time target detection in complex traffic scenes. Compared with YOLOv7, our model improves the average accuracy by 3.7% on the SODA 10M dataset, and the mean average precision (mAP) value reaches 75.9%.目标检测是人工智能发展的核心方向之一，亦是实现智能车辆与高级驾驶辅助系统环境感知的关键环节。然而，当前应用于复杂真实场景的目标检测算法仍存在诸多棘手难题，例如适配多样道路场景能力不足、实时检测性能欠佳等。因此，亟需平衡目标检测器的运算效率与检测效能。YOLOv7作为单阶段目标检测算法，集成了多项先进模块与技术，旨在实现更精准、更快速的目标检测。本文以YOLOv7为基准模型，结合可变形卷积（deformable convolution）与注意力机制模块（attention mechanism module），实现复杂场景下的充分检测，并平衡了计算效率与检测精度。此外，通过融合轻量级网络模块，在提升检测器特征表达能力的同时加快了模型的运算速度，进而完成复杂交通场景下的实时目标检测任务。相较于YOLOv7，我们的模型在SODA 10M数据集上的平均精度提升了3.7%，平均精度均值（mean average precision，mAP）达到75.9%。