Human Activity Recognition by Tracking the Global Motion.

Date of Submission

December 2013

Date of Award

Winter 12-12-2014

Institute Name (Publisher)

Indian Statistical Institute

Document Type

Master's Dissertation

Degree Name

Master of Technology

Subject Name

Computer Science

Department

Electronics and Communication Sciences Unit (ECSU-Kolkata)

Supervisor

Mukherjee, Dipti Prasad (ECSU-Kolkata; ISI)

Abstract (Summary of the Work)

Human Activity Recognition is an active area of research in computer vision with wide scale applications in video surveillance, motion analysis, virtual reality interfaces, robot navigation and recognition, video indexing, browsing,etc. It consists of analyzing the characteristic features of various human actions and classifying them.In a video with static background, activity analysis generally consists of foreground detection, forming the human trajectory ,feature selection and then classifier. However, in real world situation the assumption of static background does not always hold.Learning global motion patterns from a video is an activity classification problem is important, especially in noisy environment where there is illumination changes,jitters,camera motion and also background is not same for all the videos used for classification.In our approach we define a method to tackle a real world situation containing illumination changes,jitters and camera motion as an inherited noise in the system. Also a given activity is being performed under different backgrounds. We compute the dense optical flow and quantize it into different labels.Then correct the alignment of the optical flow vectors using probabilistic relaxation labeling in each frame and along the time axis to achieve the dominant motion. We only retain the processed optical flow vectors which are locally maxima.These step removes some amount of noise in the video.It is followed by the construction of the tracks which is a sequence of 3 Dimensional points based on the dominant motion of the system representing the activity.These tracks representing the global motion of the system are not much effected by the induced noise in the videos.We select top dominant tracks of the system based on a criterion which is further processed to represent as the feature descriptor of the given activity.The efficacy of the approach is demonstrated on challenging LIRIS dataset.

Comments

ProQuest Collection ID: http://gateway.proquest.com/openurl?url_ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri:pqm&rft_dat=xri:pqdiss:28843121

Control Number

ISI-DISS-2013-286

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

DOI

http://dspace.isical.ac.in:8080/jspui/handle/10263/6442

This document is currently not available here.

Share

COinS