Multi view based human action recognition and behavior understanding using shape features

Abstract

Humans have the ability to recognize an event from a single still image. It is the natural tendency of the human beings to give more attention to dynamic objects compared to static objects in a scene. Human motion analysis is currently one of the most active research topics in computer vision. It has three tasks, namely detection or low level processing, mid level processing or tracking and high level vision or recognition. There is a need for automated human action recognition (HAR) system which could recognize human actions and subsequently to analyze their behavior in order to understand their motion patterns. This thesis proposes orientation information based shape representation scheme with reduced dimensionality and computational complexity. The proposed orientation context based shape features, namely Triangulated Shape Orientation Context (TSOC) and Centroid Orientation Context (COC) are proposed which are robust to noise, occlusion and multiple view points. The viewpoint invariance helps to learn and recognize human shape using different camera configurations. Consideration of outer periphery boundary pixels during feature extraction provides invariance to different clothing styles. Similarly, for ViHASi test data set with 20 actions and 72 samples per action, this work has achieved an average 95.42% accuracy of recognition. The results demonstrate suitability of the proposed work for real world practice. The comparative analysis reported ensures the reliability of the multi-view based human action recognition using TSOC and COC shape features. Currently, the vision based analysis for human behavior is performed by assigning multiple analysts to watch the same video stream continuously. The proposed work achieved 86% specificity and 84% sensitivity for training data set. For the test data set, the system achieved 85% specificity and 84% sensitivity. This proposed model could be effectively utilized to learn the real world scenarios where behavior understanding is a complex task. newline newline newline