Pedestrian Walk Direction Estimation for Smart Surveillance
Rahul Raman · 2019
Study of human behavior during a walk is crucial for surveillance, traffic control, assisted living, crowd flow in shopping malls, and various other domains. In this respect, direction of walk estimation is an imperative research issue. The challenge of pedestrian walk direction estimation needs to be addressed at different working environment, with different camera setup, at different light and noise levels, for different crowd types, and with different objectives. Hence, the challenges are multifold. This thesis presents an extensive study on different research works performed in last couple of decades towards pedestrian walk direction estimation, comparing their work at both quantitative and qualitative levels. The thesis also performs an elaborated study on existing databases for visual surveillance and proposes a suitable database for walk direction estimation. The thesis further proposes machine learning based classification of pedestrian walk direction among eight pre-identified equiangular discrete directions. Exploiting feature set from temporal domain, proposed method eradicates the low accuracy of classification among directions directly towards and away from camera, thus yielding overall higher accuracy of walk direction estimation. The thesis further proposes a kinesiology inspired spatial feature to be used along with temporal feature to yield a robust feature set originated from kinematics of pedestrian locomotion along sagittal plane. Overall pedestrian direction estimation result are more precise and gives correct result with sudden change in walk direction or even orientation of pedestrian to predict probable direction of motion. These results are generated using hidden Markov model (HMM) and Least Square Support Vector Machine (LS-SVM) methods with sufficient theoretic and experimental justification. The demand of higher accuracy of direction estimation result and not-so-discrete nature of pedestrian motion has further inspired to propose the fuzzy approach based direction estimation to yield direction of pedestrian motion by defuzzification of spatial feature. The proposed analysis and implementation of type-1 fuzzy based direction estimation over same feature set gives crisp angular direction (in degree) with respect to view axis. Proposed different levels of direction estimation suits different traffic, surveillance, and research related needs to counter different noise levesl of pedestrian video input and different levels of required precision for direction estimation result. Proposed direction estimation method is robust to partial occlusion, improper segmentation and can give discrete classification result and even beyond discrete levels which may be required for multiple applications.