Sparse Tensor-based Point Cloud Attribute Compression

Abstract

Surveillance videos can capture a variety of realistic events and also anomalies. Due to an increase in the crime rate in public areas, surveillance cameras are adopted in a very large number. But as these crimes/public disputes are rare to occur at a specific location, human monitors are idle most of the time. Hence, there is a justified need to develop intelligent systems for anomaly detection. There are several seminal deepneural architectures proposed in this field of anomaly detection ranging from using deep learning as a feature extraction tool to complete end-to-end deep-learning-based anomaly detection models. Any practical anomaly detection model must be generic in detecting a spectrum of anomalous events; however, several models can detect only specific types of anomalies. Further, several models are not amenable to distributed training over many machines on large streaming data, which is typical in a video surveillance system. In this paper, we discuss the techniques to detect anomalies in real-time by exploring recent architectures in the literature and analyze and explore ways we can improve the detection accuracy of the model. We propose a batching methodology that improves the existing model's area under the curve by 2%.