Artificial Intelligence Based Efficient Activity Recognition with Real Time Implementation for ATM Security

Recognizing human activities plays a substantial role in human-to-human and human-to-computer interactions. Recognizing human activities from video sequences or pictures is a difficult task because of troubles, such as history clutter, partial occlusion, modifications in scale, viewpoint, lights and look. Human action is difficult to classify as a time series. Predicting a person’s movements is a part of this. In this paper, the KTH video dataset is used for designing the system. Feature extraction methods like optical flow and spatiotemporal techniques are being utilized to extract the features. Triple stacked autoencoders are used for clusterization to reduce the data dimensions. An efficient BoW vector feature extraction method is used for extracting text data, by which data is obtained for training the model. A deep learning algorithm such as VGG19 is used to determine and classify the activities of a human. The objective of this efficient model is to apply as an ATM surveillance as a camera module fixed in the room to perform constant surveillance. The Police department can have an mobile application through which they can monitor and desist any unwanted human activities happening in the ATM.