Research On Hierarchical Encoding Of Motion Perception And Its Application

Visual feature,which is widely adopted to tackle classification and other various visual tasks,represents the intrinsic characteristics of different things and its advantages or disadvantages affects the performance of models directly.The problem that how to characterize motion information efficiently to achieve accurate action detection or recognition has always attracted increasingly researchers' attentions to develop skills to analyze motion patterns in video scenes.Moreover,the adoption of hand-crafted features is a clear limitation of previous methods,as it enforces some prior knowledge that is very difficult to define especially in complex videos.Although deep learning architecture inspired from biology owns better capability of encoding,this framework is merely trained as an end-to-end system with a ‘black-box' mode,ignoring the potential characteristics within hierarchical encoding of perception in terms of motion patterns.In addition,it also lacks of heuristic knowledge in each level to design well-suited deep network for activity descriptors.Feature extraction,as one of the problems in anomaly detection task,is mainly reflected in how to represent motion information efficiently and model the usual patterns.Anomaly detection,which is a topic of great interest in video scene analysis,mainly describes the dynamics that either occur occasionally or deviate from the ordinary patterns.Moreover,the other challenge of this task is summarized as follows: how to estimate behaviors as abnormal ones accurately.Similarly,the complexity and diversity of scenes also make anomaly detection still a very challenging vision task.To solve the above-mentioned problems,the main work of this dissertation is summarized as follows:(1)For the issue of perception of motion encoding,the hierarchical learning model that conforms to the regularity exists in Human Vision System is formed by utilizing the unsupervised learning algorithm.Based on video scenes,the underlying principles within hierarchical encoding in terms of motion patterns are researched,and then they are finally described as universal theories.(2)In view of the problems of encoding and estimation principles in abovementioned anomaly detection task,the model of anomaly detection with the hierarchical learning framework is learned by taking advantage of the encoding regularities and considering appearance and motion representations of dynamic regions.Then,ordinary motion patterns are explored from joint high-level features in unsupervised way,and their distribution is also estimated through non-parametric kernel density approach.Finally,the one-class SVM model is utilized to predict the anomaly score of each pattern.(3)Due to the shortcomings of the existing action proposal methods in terms of the representations of visual hierarchical perception,a biological HMAX-like model for proposing action candidates is presented to locate spatial action regions in video frames and then link the predictions to produce consistent detections in time,named as action tubes.In addition,the influence of this proposal model on anomaly detection is also explored and discussed.