基于多网络级联预测的异常行为识别方法研究

Research on Abnormal Behavior Recognition Method Based on Multi-Network Cascade Forecast

异常行为识别是通过计算机提取图像序列中的特征信息,构建行为规则模型,实现对异常行为的分类和识别.现有端到端的基于深度学习的异常行为识别方法,受数据集种类和规模影响,模型自适应能力差,对人类行为的刻画能力有限.而且,异常行为定义一般取决于场景,精确分类困难.为了实现对多人员异常行为快速识别,结合行人在正常运动具有规律性,且异常事件和违规操作具有连续性,提出了一种基于多网络级联预测的异常行为识别方法.采用实例分割模型提取人体位置,利用提取骨架与稀疏光流相结合,完成视频中骨骼的跟踪,通过双向递归编解码网络,预测动态骨架信息,将骨架异常分数与阈值对比,判断行为异常.最后通过ShanghaiTech Campus公开数据集和自制数据集进行测试,实验结果表明,本文方法在不同场景、不同异常行为下都有较高的检测精度.

Abnormal behavior recognition is to extract the characteristic information in the image sequence through the computer.A behavior rule model is built based on feature information.The model can classify and identify abnormal behaviors.The existing end-to-end deep learning-based abnormal behavior recognition method is affected by the type and scale of the data set.The model has poor adaptive ability and limited ability to portray human behavior.The definition of abnormal behavior generally depends on the scene,it's difficult to achieve accurate classification.In order to realize the rapid recognition of abnormal behaviors of multiple people,combined with the regularity of pedestrians in normal movement,and the continuity of abnormal events and illegal operations,an abnormal behavior recognition method is proposed based on multi-network cascade prediction.The instance segmentation model is used to extract the position of the human body,and the skeleton is combined with the sparse optical flow to complete the tracking of the skeleton in the video.The dynamic skeleton information is predicted through the two-way recursive codec network,and the skeleton abnormality score is compared with the threshold to judge abnormal behavior.The proposed method is tested on both the public dataset and the self-made dataset of Shanghaitech Campus,and the experimental results show that the proposed method has high detection accuracy under different scenarios and different abnormal behaviors.