基于3D-LCRN视频异常行为识别方法

3D-LCRN based Video Abnormal Behavior Recognition

摘要:

自动准确识别监控视频中的异常行为在安防领域具有广泛的应用前景.本文提出一种基于3D-LCRN(3D Long-short-term Convolutional Recurrent Network)视觉时序模型的视频异常行为识别方法.首先,基于视频图像帧间的结构相似性,结合光照感应与光照补偿机制进行背景建模,获取对光照突变与背景运动不敏感的矫正光流场与矫正运动历史图.同时,针对异常与正常行为视频数据失衡问题,计算三通道矫正光流运动历史图COFMHI(corrected optical flow motion history image),随机提取视觉词块进行聚类,对样本数量与维度进行双向扩充,充分获取样本的微分和积分运动信息.在此基础上,采用3D-CNN深度学习网络模型对COFMHI进行学习,获取局部短时序时空-域特征,结合可学习贡献因子加权的LSTM网络以压制无关、冗余、具有混淆性的视频片段,进一步提取由短时序-长时序,由局部-全局的多层次时-空域特征用于异常行为识别.通过与同类方法的客观定量对比,实验结果表明,本文方法在光照突变与背景运动等复杂场景下具有优异的异常行为识别性能,进一步表明该方法有效、可行.

Abstract:

Automatically anomaly recognition in surveillance videos is a crucial issue for social security. A 3D-LCRN visual time series model was proposed for abnormal behavior recognition on video surveillance. Firstly, a structural similarity background modeling method was proposed to obtain corrected optical flow and corrected motion history image, which was insensitive to illumination variation and background moving against background interference in complex scenes. Secondly, a new sample expansion method was proposed to solve the imbalance between normal training samples and abnormal ones, which enriched the spatial and temporal information of samples from both dimensionality and quantity. On dimensionality, the method stacked corrected optical flow and corrected motion history image to generate the corrected optical flow motion history image. In quantity, COFMHI was randomly cropped and clustered into center visual words by K-means. Finally, COFMHI was used as 3D-CNN input to extract local short-time spatial-temporal features of behavior. In order to suppress irrelevant, redundant and confusing video clips, a learnable contribution factor weighted LSTM was used to deeply extract the global long-time spatial-temporal features for abnormal behavior recognition. Through 3D-LCRN, abundant spatial-temporal features were extracted from both local to global and short-time to long-time levels. Experimental results show that the proposed method has excellent performance of abnormal behavior recognition in complex scenes such as illumination variation and background moving in comparison with the state-of-art methods.