Identification of undesirable behaviour in CCTV footage using Deep Learning

Anomaly detection in CCTV recording is a difficult and challenging subject due to the issue of the vast amount of the data that must be processed, and the expertise required to analyse it. CCTV operators undergo a long and extensive training to spot anomalous behaviour in CCTV recording, but even with the acquired expertise, on average an operator will lose up to 45% of screen activities after 12 minutes, and up to 95% after 22 minutes.
This research investigates a novel pipeline technique to process CCTV recording using a combination of different unsupervised machine learning techniques. The principle pipeline technique evaluated consists of and Autoencoder as a feature extractor, in combination with a one-class Support Vector Machine (SVM), and Hidden Markov Model (HMM). Extracted Autoencoder features are categorised using the SVM to determine anomaly per frame, followed by temporal smoothing of the SVM frame categorisation with the HMM. The system achieves an accuracy of 61.38% and an AUC of 0.59.
The system was evaluated by comparing the results produced by the system with regards to labels provided with a dataset. The results collected from the comparison were used to produce an area under curve value.
The report will look in to comparing the results of using a pre-trained CNN (VGG16) and Autoencoder for purpose of feature extraction.
Being unsupervised, the system requires very little human interference and it was designed to teach itself how to differentiate an anomaly from a non-anomalous event. The only human input that the system requires was the selection of parameters for all the algorithms. The rest was left for algorithm to decide based on a set criterion. The obtained results, which while inevitably inferior to the performance of comparable supervised systems (i.e. where the anomaly class is explicitly labelled in the training data), provides an effective proof of concept of pipelining that can be used for purpose of unsupervised anomaly detection of a CCTV image frame.