Extreme learning machine based tracking-learning-detection framework in surveillance video

• Main Author: Lai, Tinglin
• Other Authors: Huang Guangbin
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/67411
• Institution: Nanyang Technological University

The importance of surveillance videos is gaining more attention from companies and organization nowadays. As one of the important methodologies in surveillance video and computer vision domain, object tracking emerged as a trending topic. Numerous researches have been done to further explore this topic. In this project of seeking object tracking algorithm optimisation, the author went through the full process of the research including data collection, literature review, implementation, evaluation and modification. In the data collection part, the author compared different datasets including public datasets, sample videos from Delta Company and YouTube videos. After thorough examination, dataset “visual tracking benchmark” has been selected for the later evaluation process. In the implementation process part, the author got a detailed research and understanding of the algorithm and source code for the TLD framework. Detailed comments for the most important two files are illustrated in the report. In the evaluation part, visual tracking benchmark datasets are used for evaluation including two indicators: numbers of the successfully tracked frames and average localisation errors. Both advantages and shortcomings are identified in this section. In the modification part, based on the secondary research, the author implemented an ELM (Extreme Learning Machine) based TLD framework which replaced the original k-NN classifier with extreme learning machine classifier to achieve better classification accuracy and smaller standard derivation. Differences in performance of the original and new algorithms are evaluated using the two indicators mentioned above. Lastly, some head-ups are recommended by the author as a guide for future researchers to improve the current algorithm and implementation.