Models of human movement detection in video frames using machine learning technology

Publication Type:
Thesis
Issue Date:
2019
Full metadata record
In recent years, human movement detection has been a very active and vibrant field of research. Examples of successful applications that adopt discoveries and developments in the human movement detection domain include pedestrian detection, intelligent monitoring systems and activity recognition. Thus far, much research work has been done and many different approaches explored to optimise accuracy, performance and productivity of human detection system. However, there are still many open questions related to these issues remain unresolved. The following study aims to develop a more effective human detection system that is able to operate in various environmental conditions and application contexts including illumination changes, pose and scale differences. The existing solutions for human detection and tracking systems often do not produce reliable and consistent results without considering changes in environmental conditions. In this work, a novel illumination invariant human detection algorithm is proposed which applies an alternative approach for the selection of orientation extraction and texture extraction features to identify human shapes in various illumination and contrast invariant conditions. An innovative human detection approach is also proposed to resolve and improve results of pose invariant cases. This research work involves the exploration of feature extraction techniques that offer superior results when dealing with human subjects in pose invariant conditions. Another innovation of this investigation is the design of a human detection and tracking model that can work in situations where human subjects are occluded within frames. In the proposed models, several pre-processing and post-processing stages are used for reducing detection errors and to improve the model performance. These approaches help to classify the frame contents more efficiently. The proposed computational solutions are extensively tested and performance evaluated using the standard datasets. The resulting output is encouraging when compared to the reported and the State-Of-The-Art human detection algorithms. The newly developed methods are tested using two practical applications and are included in this thesis as action research studies. In the first action study, children activity monitoring system is built to test the human movement detection algorithms, whilst the second action study involves a construction of an expert system for counting humans in a moving crowd to validate the effectiveness of the proposal computational models. These two key action research studies report high feasibility and viability of the proposed solutions
Please use this identifier to cite or link to this item: