改进Faster-RCNN的工程车辆识别方法

Improved Faster-RCNN Method for Engineering Vehicle Detection

针对Faster-RCNN中预设锚点框尺寸与数据集不匹配、RoI Pooling中量化操作带来的区域不匹配问题,研究了一种改进Faster-RCNN的方法,实现城镇视频监控场景下工程车辆(铲车、挖掘机及大货车等)的快速、准确识别。首先,使用K-means算法对训练数据中的目标边框聚类分析,通过聚类分析得到7种较优的锚点框,将聚类结果用于区域建议网络;其次,以区域特征聚集方法RoI Align替换ROI Pooling,解决因量化操作带来的区域不匹配问题;实验结果表明改进Faster RCNN方法的精度更高,mAP达到87.73%,提高14.06%。

Aiming at the problem of the mismatch between the preset anchor boxes size in Faster-RCNN and the target boxes in dataset, and the regional mismatch caused by the quantization operation in RoI Pooling, an improved method based on Faster-RCNN is studied to realize rapid and accurate identification of engineering vehicles in urban video surveillance scene. First, we clustered the target boxes in the training data by K-means and obtained seven appropriate anchor boxes, which was applied in RPN. Then, substitute RoI Align for RoI pooling in Faster RCNN. RoI Align can effectively solve the limitation of region mismatch caused by quantization operation in RoI pooling. Finally, the experimental result shows that the improved method has a higher accuracy, and the mAP reaches 87.73%,an increase of 14.06%.