Criminal activity detection from videos under low light condition using deep neural network

Abstract

Criminal activity detection from footage, especially in low-light circumstances, offers a considerable problem due to reduced visibility, noise, and detail loss. In this research, we present an approach for detecting criminal actions in low-light videos using deep learning models. The specific difficulty of low-light conditions, which result in limited visibility and noisy data, is handled using modern video pre-processing techniques to improve video quality. Our strategy improves video classification accuracy by using both spatial and temporal information, preserving essential visual signals. We use transfer learning to adapt pre-trained models such as VGG-16, MobileNetV2, NASNetMobile, LSTM and a Conv-LSTM so that they can generalize effectively in low-light settings. The modified UCF Crime dataset simulates low-light situations, and our results illustrate that the suggested technique improves detection accuracy while remaining efficient. This study paves the path for more dependable surveillance systems capable of working in harsh environments. After working through all these models the best performing model we got was the ConvLSTM model where we got an accuracy of 77%.