Crime Alert Through Smart Surveillance Using Deep Learning Techniques
Aarushi Dua, Bhavya Kalra, Abhishek Bhatia, M. S. Madan, Anuradha Dhull, Yogita Gigras · 2022
In this paper, we proposed a way to detect different categories of abnormal activities happening in India such as fight, explosion and shootout. The goal is to build a model that can identify signs of violence and aggression in videos and separates out anomalies from normal patterns. The problem statement of crime detection is achieved in two broad steps: One by building a deep learning model (ConvLSTM) to categorize different crimes and then deploying this model to an interface where live stream footage is connected to the server for any crime alert. The model is evaluated on UCF-crime dataset and if any frame in live stream captures a criminal act, the tool will issue a detection warning for a danger situation, signaling suspicious actions at a certain point in time. The suggested ConvLSTM2D approach outperforms the traditional convolutional neural networks-long short-term memory (CNN-LSTM) algorithms in terms of accuracy. The performance evaluation of training data is based on Area under the Receiver Operator Characteristic (ROC) curve.