基于运动外观多通道层级ICA编码的异常检测

Hierarchical ICA Encoding Combined with Motion-Appearance Information for Anomaly Detection

摘要: 针对现有异常表示方法对视觉感知层级关系描述能力的不足，基于生物视觉感知编码特性启发，本文提出一种基于运动外观多通道层级ICA编码模型，实现复杂场景中的异常检测任务。首先，对现有的生物视觉层级编码框架，进行三级逐层学习拓展，采用ICA统计方法提取层内视觉感知编码模式，利用HMAX机制实现层级信息传递。其次，借助视觉双通道处理机制，各通道独立完成三层编码模式学习，随后联合双通道特征构建异常模式表达，最终，利用单类支持向量机模型对正常和异常情况进行判定。在UCSD数据集上，分别验证了本文方法的运动感知编码特性和异常检测的性能，实验结果能够说明本文异常模式表达优于现有的手工设计特征，以及深度学习特征。

Abstract: Due to the shortcomings of the existing anomaly detection methods in terms of the representation of visual hierarchical perception, inspired by the regularities in perception encoding of bio-mimetic vision, a hierarchical ICA encoding approach combined with motion-appearance is presented for abnormal events detection. First of all, this method extends the Existing biological visual hierarchy coding framework with three-level layer-wise learning, and uses ICA statistical method to extract intra-layer visual perceptual coding patterns, and utilizes the HMAX mechanism to transmit the Hierarchical information. In addition, with the processing theories of double channels in the visual system, Each channel is proposed to separately complete three-layer coding pattern learning. Then, these double features are fused to represent the Anomaly patterns. Based on the joint representations, the one-class SVM model is used to predict the abnormal score for each input. The proposed method, whose properties of motion perceptual coding and the performance of anomaly detection, is evaluated on UCSD datasets, and the results demonstrate that the learned feature representations in this paper for anomalous patterns are superior to other traditional hand-crafted features and deep learning features.