Abnormal Activity Detection Using Deep Learning

Abstract

Identifying abnormal activity is a laborious job, and this has led to the advancement in the domain of deep learning for surveillance which assured performance gain. Abnormal detection plays a pivotal role in the field of research and application systems. Detecting real-world problems such as accidents, burglary, explosion, fighting, robbery and other critical events is crucial, so we developed a deep learning-based algorithm to reduce manual work and time. To provide an optimal solution for the same, we have designed a model that detects abnormal events. To achieve this objective, we have implemented two approaches. The first approach is the model that we have trained (auto-encoder), and the second approach is the pre-trained model (C3D feature extraction). In approach 1, we have introduced a spatiotemporal auto-encoder, which is based on a 3D convolution neural network. After training the model with auto-encoder, our model has reconstructed the frames with a minimum loss of 8% and a maximum accuracy of 71%. To improve accuracy and to reduce loss, we have exercised approach 2 (C3D extraction) that computes features from the fully connected layers of the C3D which resulted in increased accuracy. To detect the abnormal activity in banks, roads and many more crowded areas, this can be implemented through CCTV cameras to automate and simplify manual work.