Privacy-Aware Anomaly Detection Using Semantic Segmentation

This paper sets out to investigate if semantic segmentation can be used to achieve anonymity in video surveillance while maintaining the ability to perform anomaly detection. The paper is centered around finding the best model for segmenting the anomaly detection dataset UCHK Avenue without semantic ground truth available. To do this, a series of segmentation models pre-trained on ADE2OK and Cityscapes are evaluated against a custom semantic annotation of selected frames from Avenue and the segmentation results are compared both quantitatively and qualitatively. The segmented dataset is then tested on a series of different anomaly detection baselines and the results are compared both in terms of global and anomaly specific accuracy. When comparing the anomaly detection accuracy for RGB and segmented data it was found that anonymity in anomaly detection can be achieved at a small cost in global accuracy but with better accuracy for some specific anomalies.