Real-time Anomaly Detection in Surveillance Feeds

Abstract: Rapid advances in the surveillance infrastructure have enabled us to capture normal and anomalous events at scale, coupled with tremendous progress in computer vision and pattern recognition. However, the issue of timely response to potential threatening situations is still a problem at large. Various challenges such as low quality feeds, occlusion , clutter, lack of training data, adversarial attacks make it extremely hard for the network to achieve the desired and timely accuracy and performance, leading to hazardous situations that could have been potentially avoided. In this paper, we study state of the art approaches to tackle this problem and study their capabilities and limitations. Furthermore we also present the results of several experiments conducted to tackle this challenge from a supervised, unsupervised, generative and reinforcement perspective. We hope to present these results as an enabler for future work in this area.

Bio: Utkarsh Contractor is the Director of AI at Aisera, where he leads the data science team working on machine learning and artificial intelligence applications in the fields of Natural Language Processing and Vision. He is also pursuing his graduate degree at Stanford University, focussing his research and experiments on computer vision, using CNNs to analyze surveillance scene imagery and footages. Utkarsh has a decade of industry experience in Information Retrieval and Machine Learning working at companies such as LinkedIn and AT&T Labs.