Marker-less human body part detection, labelling and tracking for human activity recognition

Preview

PDF WRAP_Thesis_Azhar_2015.pdf - Requires a PDF viewer. Download (39MB) | Preview

Request Changes to record.

Abstract

This thesis focuses on the development of a real-time and cost effective marker-less computer vision method for significant body point or part detection (i.e., the head, arm, shoulder, knee, and feet), labelling and tracking, and its application to activity recognition. This work comprises of three parts: significantbody point detection and labelling, significant body point tracking, and activity recognition. Implicit body models are proposed based on human anthropometry, kinesiology, and human vision inspired criteria to detect and label significant body points. The key idea of the proposed method is to fit the knowledge from the implicit body models rather than fitting the predefined models in order to detect and label significant body points. The advantages of this method are that it does not require manual annotation, an explicit fitting procedure, and a training (learning) phase, and it is applicable to humans with different anthropometric proportions. The experimental results show that the proposed method robustly detects and labels significant body points in various activities of two different (low and high) resolution data sets. Furthermore, a Particle Filter with memory and feedback is proposed that combines temporal information of the previous observation and estimation with feedback to track significant body points in occlusion. In addition, in order to overcome the problem presented by the most occluded body part, i.e., the arm, a Motion Flow method is proposed. This method considers the human arm as a pendulum attached to the shoulder joint and defines conjectures to track the arm since it is the most occluded body part. The former method is invoked as default and the latter is used as per a user's choice. The experimental results show that the two proposed methods, i.e., Particle Filter and Motion Flow methods, robustly track significant body points in various activities of the above-mentioned two data sets and also enhance the performance of significant body point detection. A hierarchical relaxed partitioning system is then proposed that employs features extracted from the significant body points for activity recognition when multiple overlaps exist in the feature space. The working principle of the proposed method is based on the relaxed hierarchy (postpone uncertain decisions) and hierarchical strategy (group similar or confusing classes) while partitioning each class at different levels of the hierarchy. The advantages of the proposed method lie in its real-time speed, ease of implementation and extension, and non-intensive training. The experimental results show that it acquires valuable features and outperforms relevant state-of-the-art methods while comparable to other methods, i.e., the holistic and local feature approaches. In this context, the contribution of this thesis is three-fold:

Pioneering a method for automated human body part detection and labelling.

Developing methods for tracking human body parts in occlusion.

Designing a method for robust and efficient human action recognition.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item