基于改进YOLOv7的高效行人检测方法

Efficient pedestrian detection method based on improved YOLOv7

针对目前的行人检测方法无法在复杂环境下同时满足高准确率和高检测速度的问题,提出了基于改进YOLOv7(You Only Look Once version 7)的高效行人检测方法。首先,通过鬼影混洗卷积(GSConv)与VoVGSCSP(VoVNetGS Conv Cross StagePartial)构建Slim-Neck,前者使用混洗操作将普通卷积生成的信息渗透到可分离卷积的输出中,来实现通道间信息的交互,后者采用一次聚合方法设计了跨阶段部分网络,VoVGSCSP模块降低了计算量和网络结构的复杂性,并保持了足够的精度;其次,在YOLOv7输出部分引入卷积注意力模块(CBAM),利用通道注意力和空间注意力来捕获特征之间的相关性,从而优化YOLOv7的特征表示能力,提高方法的准确性和鲁棒性。实验结果表明:在多个行人数据集上,与YOLOv5和YOLOv7相比,改进的YOLOv7方法平均精度(AP)提升了1.63~3.51个百分点,对数平均缺失率(LAMR)降低了0.54~3.97个百分点;相较于YOLOv7平均检测速度提升10FPS;同时通过弗里德曼检验结果证实改进的YOLOv7方法可用于实际数据,有效地实现了复杂环境下高精度、快速的行人检测。

Aiming at the problem that the current pedestrian detection methods cannot satisfy both high accuracy and high detection speed in complex environments,an efficient pedestrian detection method based on improved YOLOv7(You Only Look Once version 7)was proposed.Firstly,Slim-Neck was constructed through Ghost-Shuffle Conv(GSConv)and VoVGSCSP(VoVNet GSConv Cross Stage Partial).In GSConv,shuffling operations were employed to merge information from standard convolution into separable convolution output,realizing channel-level interaction;in VoVGSCSP,a one-time aggregation approach was employed to design a cross-stage partial network,reducing network complexity while maintaining high accuracy.Secondly,the Convolutional Block Attention Module(CBAM)was introduced in the output part of YOLOv7,in which channel attention and spatial attention were used to capture the correlation between features,thereby optimizing the feature representation ability of YOLOv7 and improving the accuracy and robustness.The experimental results showed that:on multiple pedestrian datasets,the proposed method increased the Average Precision(AP)by 1.63 to 3.51 percentage points,reduced the Log Average Miss Rate(LAMR)by 0.54 to 3.97 percentage points,compared with YOLOv5(You Only Look Once version 5)and YOLOv7;and enhanced the average detection speed by 10 FPS(Frames Per Second)compared to YOLOv7.The Friedman test results verify that the proposed method can be applied to actual data well,and effectively achieves high-precision and fast pedestrian detection in complex environments.