Deep Learning and Vision-Based Systems for Crime Detection and Prevention in Urban Surveillance

This paper presents the training and implementation of machine learning and computer vision algorithms aimed at the real-time prevention and detection of criminal activity. The proposed approach is motivated by the increasing crime rates in Ecuador and seeks to reduce police response times through the automated analysis of suspicious behaviors in video footage. Deep neural network models were employed, specifically the I3D architecture and a hybrid ISD+ConvLSTM model, capable of identifying anomalous patterns in temporal video sequences. Experimental results demonstrate the effectiveness of these techniques in both anticipating and detecting criminal events, thereby providing a valuable tool for enhancing public safety. This work represents a significant advancement in the application of computer vision to urban environments and offers a practical solution with strong potential for real-world deployment.