Anomalous behavior detection using Spatio temporal Feature and 3D CNN model for Surveillance

Abstract

Anomalous and violent action detection has become an increasingly relevant topic and active research domain of computer vision and video processing, within the past few years. It has many proposed solutions by the researchers and this field attracted new researchers to contribute in this domain. Furthermore , the widespread use of cameras used for security purposes in big modern cities has also allowed researchers to research and examine a vast amount of information so that autonomous monitor ing can be executed. Adding effective automated violence unearthing to videotape security or multimedia content watching technologies (CCTV) would make the task of carpoolers, walk organizations, and those who are in control of social media activity monitoring much easier. We present a new deep scholarship skeleton for determining whether a videotape is violent or not, based on a suited version of DenseNet , and a bidirectional convolutional LSTM module that allows unscram bling pointed Spatio-temporal features in this paper. In addition, ablation research of the input frames was carried out, comparing thick optic outpouring and touching frames. Throughout the paper, we analyze various strategies to detect violence and their classification in use. Furthermore, in this paper, we detect violence using the Spatio-temporal feature with 3D CNN which is a DL violence detection framework, specially better for crowded places. Finally, we used embedded devices like Jetson Nano to feed with dataset and test our model and evaluate. We want a warning sent to the local police station or security agency as soon as a violent activity is detected so that urgent preventive measures can be taken. We have worked with various benchmark datasets where in one dataset, multiple models achieved a test accuracy of 100 percent, making them invincible. Furthermore, for a different dataset our models have shown 99.50% and 97.50% accuracy rates. We also did a cross dataset experiment in models which also showed pretty good results of higher than 60%. The overall results we got suggests that our system has a viable solution to anomalous behavior detection.