计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (20): 207-214.DOI: 10.3778/j.issn.1002-8331.2406-0204

• 图形图像处理 • 上一篇    下一篇

Haar小波下采样优化YOLOv9的道路车辆和行人检测

李琳,靳志鑫,俞晓磊,王安红   

  1. 1.太原科技大学 电子信息工程学院,太原 030027
    2.太原科技大学 体育学院,太原 030027
    3.江苏省质量和标准化研究院 国家射频识别产品检验检测中心(江苏),南京 210029
  • 出版日期:2024-10-15 发布日期:2024-10-15

Road Vehicle and Pedestrian Detection Based on YOLOv9 for Haar Wavelet Downsampling

LI Lin, JIN Zhixin, YU Xiaolei, WANG Anhong   

  1. 1.School of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030027, China
    2.School of Physical Education and Sports, Taiyuan University of Science and Technology, Taiyuan 030027, China
    3.National Radio Frequency Identification Product Inspection and Testing Center (Jiangsu), Jiangsu Provincial Institute of Quality and Standardization, Nanjing 210029, China
  • Online:2024-10-15 Published:2024-10-15

摘要: 在当前智能化、信息化的大背景下,为了实现无人驾驶模式复杂环境中智能收集道路的行人和车辆目标,提出了一种基于Haar小波下采样(Haar wavelet downsampling,HWD)的YOLOv9算法(HWD_YOLOv9)用于车辆与行人目标检测。Haar小波的下采样操作,降低特征图的空间分辨率,尽可能保留了边缘、纹路等细节信息,有效降低了信息的不确定性。采用交叉熵损失和广义骰子损失之和作为网络的损失函数,可以有效地度量概率分布之间的差异,且逐像素进行骰子损失计算,便于优化网络。实验结果显示,在KITTY数据集上,所提模型的平均精度均值达到了95.86%,检测帧率达到了179?FPS。与YOLOv9相比,改进后的算法能够精确地识别出复杂道路上不同尺度的车辆与行人,改善了原检测算法中的计算容量的冗余和小目标的漏检问题,为智能化的无人驾驶提供了视觉技术支持。

关键词: 小目标检测, 车辆行人, YOLOv9, 深度学习, Haar小波下采样

Abstract: In the current background of intelligence and informatization, the YOLOv9 algorithm based on Haar wavelet downsampling (HWD) is proposed for vehicle and pedestrian target detection in complex environments with autonomous driving mode to intelligently collect pedestrian and vehicle targets on the road. The operation of Haar wavelet downsampling reduces the spatial resolution of feature maps and preserves detailed information such as edges and textures as much as possible, effectively reducing the uncertainty of information. By utilizing the sum of cross entropy loss and generalized dice loss as the loss function of the network, the difference between probability distributions can be effectively measured, and dice loss calculations can be performed pixel by pixel, making it easier to optimize the network. The experimental results show that the average accuracy of the proposed model reaches 95.86%, and the detection frame rate reaches 179 FPS on the KITTY dataset. Compared with YOLOv9, the improved algorithm can accurately identify vehicles and pedestrians of different scales on complex roads, which not only improves the redundancy of computational capacity and missed detection of small targets in the original detection algorithm, but also provides visual technology support for intelligent autonomous driving.

Key words: small object detection, vehicles and pedestrians, YOLOv9, deep learning, Haar wavelet downsampling (HWD)