Application of cortical learning algorithms to movement classification towards automated video forensics

The need for proper and acceptable video forensics process is necessary due to the proliferation and advancement of video technologies in all aspects of our life, hence the legal system has heavily invested in this area. Also, the classification of the movements of objects detected from a video feed is an essential module to automate the forensic video process. Recently, new bio-inspired machine learning techniques have been proposed in the attempt of mimicking the function of the human brain. Hierarchical Temporal Memory (HTM) theory has proposed new computational learning models, Cortical Learning Algorithms (CLA), inspired from the neocortex, which offer a better understanding of how our brains function. This research aims to study the requirements of video forensic investigation and Police procedures to propose a new semi-automated post-incident analysis framework and to investigate the application of the CLA to movement classification towards an automated video forensics process. This research starts by reviewing the research related to police practices for video forensics as well as various proposed video forensic frameworks. Then a Questionnaire targeting CCTV/Video Forensics practitioners has been developed to capture the requirements for automating the video forensics process, this has been followed by proposing a new post-incident analysis framework. Then, a literature review covering state-of-the-art movement classification algorithms has been carried out. Finally, a novel CLA-based movement classification algorithm has been proposed and devised to classify the movements of moving objects in realistic video surveillance scenarios, and the test results have been evaluated. Tests applied on twenty-three videos have been conducted to detect movement anomalies in different scenarios. Additionally, in this study, the proposed algorithm has been evaluated and compared against several state-of-the-art anomaly detection algorithms. The proposed algorithm has achieved 66.29% average F-measure, with an improvement of 15.5%compared to the k-Nearest Neighbour Global Anomaly Score (kNN-GAS) algorithm. The Independent Component Analysis-Local Outlier Probability (ICA-LoOP) scored 42.75%, the Singular Value Decomposition Influence Outlier (SVD-IO) achieved 34.82%, whilst the Connectivity Based Factor algorithm (CBOF) scored 8.72%. The proposed models, which are based on HTM, have empirically portrayed positive potential and had exceeded in performance when compared to state-of-the-art algorithms