计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (24): 216-226.DOI: 10.3778/j.issn.1002-8331.2207-0203

• 图形图像处理 • 上一篇    下一篇

自适应特征融合的复杂道路场景目标检测算法

冉险生,苏山杰,陈俊豪,张之云   

  1. 重庆交通大学 机电与车辆工程学院,重庆 400074
  • 出版日期:2023-12-15 发布日期:2023-12-15

Object Detection Algorithm for Complex Road Scenes Based on Adaptive Feature Fusion

RAN Xiansheng, SU Shanjie, CHEN Junhao, ZHANG Zhiyun   

  1. School of Mechanical and Electrical and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China
  • Online:2023-12-15 Published:2023-12-15

摘要: 针对复杂道路场景下密集遮挡目标、小尺度目标检测精度低,容易出现漏检和误检的问题,以YOLOv5算法为网络基础框架,提出了一种自适应特征融合的复杂道路场景目标检测算法。引入特征融合因子,改进相邻尺度特征融合方式,增加各层网络有效样本从而提升中小尺度目标检测能力;增加浅层特征检测层,提升模型小尺度目标的学习能力;改进感受野模块,允许模型自适应选择有效感受野提取目标特征信息;引入Quality Focal Loss改善密集遮挡目标,小尺度目标的定位精度,并在特征融合网络加入注意力机制,提高算法对特征信息的有效利用。实验结果表明,相比原始算法,改进算法在公开数据集BDD100K(10类)、Udacity及自制数据集CQTransport的检测精度分别提高了6.7、4.9、7.9个百分点;在基本不降低检测速度的前提下,能较好提升复杂道路场景下的检测性能,并在一定程度上解决了检测过程中密集遮挡目标、小尺度目标出现的漏检和误检问题。

关键词: 目标检测, 复杂道路场景, 特征融合因子, 自适应感受野, 多尺度检测, YOLOv5

Abstract: Aiming at the problems of low detection accuracy of densely occluded targets and small-scale targets in complex road scenes, and prone to miss detection and false detection, a target detection algorithm based on adaptive feature fusion for the YOLOv5 algorithm is proposed. The feature fusion factor is introduced to improve the adjacent scale feature fusion method, and the effective samples of each layer of the network are increased to improve the detection ability of medium and small scale objects. The shallow feature detection layer is added to improve the learning ability of the model small scale objects. The receptive field module is improved, allowing the model to adaptively select an effective receptive field to extract target feature information. Quality Focal Loss is introduced to improve the positioning accuracy of densely occluded targets and small-scale targets, and an attention mechanism is added to the feature fusion network to improve the algorithm’s effective use of feature information. The experimental results show that, compared with the original algorithm, the detection accuracy of the improved algorithm in the public data set BDD100K(10 classes), Udacity, and the self-made data set CQTransport has been improved by 6.7, 4.9, and 7.9 percentage points respectively. It can improve the detection performance in complex road scenes, and to a certain extent solve the problem of missed detection and false detection of densely occluded targets and small-scale targets in the detection process.

Key words: target detection, complex road scenes, feature fusion factor, adaptive receptive field, multi-scale detection, YOLOv5