基于改进YOLOv3的人体行为检测

针对人体行为检测中相同行为差异大, 不同行为相似度高, 以及视觉角度、遮挡、不能实时检测等问题, 提出Hierarchical Bilinear-YOLOv3人体行为检测网络. 该网络采用YOLOv3在3个不同尺度上进行预测, 抽取YOLOv3金字塔特征提取网络中特定层作为Hierarchical Bilinear的输入, 捕获特征图的层间局部特征关系, 并在3个不同尺度上进行预测, 最后将YOLOv3和Hierarchical Bilinear两种预测结果融合. 实验结果显示, 改进后的模型相比于原网络仅增加了少量参数, 在保证检测效率的同时提高原算法的检测精度, 并在一定程度上优于当前行为检测算法.

Human Behavior Detection Based on Improved YOLOv3

This study proposes a neural network named Hierarchical Bilinear-YOLOv3 for human behavior detection due to a large disparity in the same behavior and high resemblance between different behaviors in human behavior detection, as well as problems such as visual angle, occlusion, and incapability of continuous real-time monitoring. YOLOv3 is first designed for prediction on three scales, and certain layers in its feature pyramid networks are used as inputs for Hierarchical Bilinear to capture local feature relationships between layers in the feature maps and predict the results on three scales. The integrated results of both YOLOv3 and Hierarchical Bilinear show that the improved network only adds a few parameters compared to the original one. It improves the detection accuracy of the original algorithm without lowering the detection efficiency and thus is superior to the current behavior detection algorithms.