Investigating movement detection in unedited camera footage

Abstract

Digital evidence collected from CCTVs can be of great aid in crime scene investigations. Investigators still manually review the video footage collected from a crime scene, which can be a very time consuming process, prone to human error and inefficient. The aim of this dissertation is to deliver a system that automates the process of detecting motion events within a video, so that the investigator can then analyse specifically parts of the video where the said events occur. The proposed system will also allow the investigator to modify the pre-processing and processing variables to suit their requirements, depending on the quality and type of video footage available. Additional functionalities offered by the system include filtering the motion events detected by colour and size, analysing and extracting features of regions where the motion event is detected, and finding association rules between objects that appear simultaneously in the video based on their colour. The data set that is used for evaluating the system is the Wallflower data set. Evaluation is conducted by using confusion matrices and comparing the results from the system with the true values of an oracle. The F1 Score measure is used to enable cross comparison between image sequences from the data set. The evaluation process was conducted over four different image sequences from the chosen data set, with the total number of frames evaluated amounting to 12,459 for each combination of pre-processing settings. The results obtained from the evaluation process show that the lowest and highest F1 Score across all image sequences used were those of 66.7% and 94.9% respectively. Evaluation has also shown that in order to get the best results, one must firstly identify the best pre-processing settings that best work for the specified video.