计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (12): 225-233.DOI: 10.3778/j.issn.1002-8331.2310-0197

• 图形图像处理 • 上一篇    下一篇

改进YOLOX的自动驾驶场景目标检测算法

陈商盈,倪受东,童林   

  1. 南京工业大学 机械与动力工程学院,南京 211800
  • 出版日期:2024-06-15 发布日期:2024-06-14

Improved YOLOX Algorithm for Object Detection in Autonomous Driving Scenarios

CHEN Shangying, NI Shoudong, TONG Lin   

  1. School of Mechanical and Power Engineering,Nanjing Tech University, Nanjing 211800, China
  • Online:2024-06-15 Published:2024-06-14

摘要: 针对原始YOLOX在自动驾驶场景中易出现小目标的误检及漏检问题,提出一种改进YOLOX网络模型的检测方法。增加一个检测头,显著提高了小目标的检测精度。用更快的空间金字塔池化(spatial pyramid pooling-fast,SPPF)替代传统的空间金字塔池化(spatial pyramid pooling,SPP)结构,并引入了高效通道注意力机制(efficient channel attention,ECA),以增强模型对复杂背景和小目标的识别能力,有效降低了小目标的漏检率。在网络的neck部分,采用深度可分离的卷积方法,进一步提高了计算效率。基于KITTI数据集,设计了一种新的数据增强技术,提高模型的训练稳定性。实验结果表明,优化后YOLOX算法的小目标平均精度APS和平均召回率AR分别提升了0.20和0.097,在小目标检测上取得了显著进步,并大大降低了漏检率。

关键词: 自动驾驶, YOLOX, 目标检测, 注意力机制

Abstract: Aiming at the problem that the original YOLOX is prone to false detection and missed detection of small objects in autonomous driving scenarios, a detection method based on improved YOLOX network model is proposed. Firstly, a detection head is added to improve the detection accuracy of small objects significantly. Secondly, the traditional spatial pyramid pooling (SPP) structure is replaced by spatial pyramid pooling fast (SPPF). Moreover, the efficient channel attention  (ECA) mechanismis introduced to enrich feature representation without increasing network complexity and reduce the missed detection rate of small targets effectively. In the neck portion of the network, depthwise separable convolutions are employed to further boost computational efficiency. Finally, based on the KITTI dataset, a novel data augmentation technique is designed to enhance the training stability of the model. Experimental results indicate that the APS and AR of the optimized YOLOX algorithm are increased by 0.20 and 0.097 respectively, which makes significant progress in small object detection and greatly reduces the missed detection rate.

Key words: autonomous driving, YOLOX, object detection, attention mechanism