Feature Coding In Human Action Recognition

Recognition of human action in video has increasingly attracted much attention in computer vision community, its aim is to recognize actions automatically from video or image sequence. However, it faces many realistic challenges because of background clutter, viewpoint changes, and variation of actors’ appearance. These challenges re-?ect the di?culty of obtaining a clean and discriminative video representation.Recently, Vector of Locally Aggregated Descriptors(VLAD) has shown to be a simple and e?cient coding method to obtain discriminative video representation.However, in most cases, dynamic background or camera’s motion will lead to much noise in feature extraction. The VLAD results in two issues. First, it encodes abnormal features which will slow down the coding process and get poor coding effect. Second,it doesn’t su?ciently consider the distribution of descriptor features and visual words.This paper focuses on feature coding, its contribution and innovation include:1. Fully introduction of human action recognition and related theory are given. The difference and limits of the most used feature coding methods are researched in detail.2. An improvement of VLAD, Uncertain VLAD(UVLAD), is proposed. After suf-ficiently consider the noise features and distribution of descriptor features and visual words, each descriptor feature is aggregated by multiple visual words or discarded appropriately. The performance of feature coding is enhanced.3. The e?ciency of the proposed UVLAD coding method is evaluated by experiments on KTH, You Tube and HMDB51. The experimental results are better than the state-of-art methods.