A Lightweight Approach for Detecting Crime Activities in Video Surveillance
Kamparapu V V Satya Trinadh Naidu, Narakala Sasikala, P. Sumanth, Yaleti Veeranjaneyulu, Kattela Yatham Shavali · 2025
The growing demand for public safety has made reliable crime detection in the video surveillance systems to be increasingly important. Currently, traditional surveillance methods are based on the continuous human monitoring that is both inefficient and vulnerable to errors. This has happened as a result of the increasing importance of ensuring public safety, thus the desire for a reliable crime detection in video surveillance is more important than ever. Because it's not necessarily efficient or reliable when we're relying on humans to do the monitoring all the time, we come up with traditional monitoring methods. Modern methods based on the machine learning and deep learning have been able to detect automatically, however, object and action recognition-based techniques suffer from difficulties including the requirement of high volume of labeled data and high rates of false alarms. In this study we propose a streamlined crime detection model using unsupervised autoencoder based autoencoders and joining it with the pre trained VGG16 neural network to extract the features efficiently. The model is built by focusing on the normal behavior patterns in surveillance data rather than the prelabeled crime instances. he results of the tests show strong performance where accuracy and precision are above 97% and the false positive rate is very low which means that the model is reliable. Real time application is supported by its lightweight and efficient design and is adaptable to smart surveillance systems.