SURVEILIA

Abstract

Surveillance cameras are gradually being used at every place that catches the anomalous event, yet the checking capacity of security agencies has not met the level. Traditional CCTV monitoring technologies are highly reliant on human attention to detect, infer, and monitor irregular, criminal behaviour. Moreover, the deep neural networks have limitations of resources as they are computation-intensive and training them requires a lot of training data and updated systems. To overcome this problem, our project Surveilia; automatically detects the anomaly in real-time using low resources and instantly generates the alarm in case of any suspicious behaviour by using deep learning through activity recognition algorithms and video analysis. We have trained our Deep Neural Network with two different architectures but the architecture we opted for application development is TSM+MobileNetV2 in which we achieve training accuracy of 88% with the loss of 0.3233 and test accuracy of 83% in which Abnormal class had 86% Precision and 83% Re-call whereas 86% and 88% was the Precision and Re-call of our Normal class. In the end, based on these statistics, we were able to calculate the F1-score which was an abnormal class of 85% and a normal class having 87%. We calculated FLOPs to be a performance metric for our case. The complexity for ResNet50 was 8.24 GFLOPs and for MobileNetV2 was 0.64 GFLOPs. This clearly shows that MobileNetV2 is less computationally expensive, hence, we chose it. We used the UCF-Crime dataset [1] (a video dataset) for training to extract the features through Temporal Shift Module (TSM) [2] which aims to achieve 3D-CNN accuracy by maintaining 2D-CNN complexity. It would be able to differentiate between the abnormal and normal events in the live stream. Mostly, such computations are complex and require intense computational resources. Hence, our target is to provide a deep learning model that is resource-efficient and works on simple machines such as a CPU. Furthermore, we have demonstrated our framework by porting our application (Surveilia) on different platforms i.e., Windows and Linux, and to achieve better, faster inference results we deployed Surveilia on NVIDIA Jetson Nano [3] (a small, powerful computer). The proposed project is a python-based application developed using OpenCV, PyQt5, PyTorch, NumPy, SQLite3, and NVIDIA’s CUDA.