Convolutional Neural Network-Based Theft Detection Using PYNQ-Z2 Board
Aditi S Aravinda, Anitriya Chakraborty, Disha Yadav J, Inchara P Bhat, Vipula Singh · 2025
Theft detection in surveillance systems is an important challenge, specifically in situations where manual monitoring is not always possible. This research proposes a Convolutional Neural Network (CNN)-based approach to automate detection of theft using individual image frames extracted from surveillance videos. The model aims to classify images as either "theft" or "no-theft" by examining the visual features present in each frame. Its preprocessing pipeline helps the model focus on important visual clues that are linked to theft. After training and testing with a labeled dataset, it achieved accuracy 75%. This shows that it can find unusual activities. Performance of the system was assessed by using both Receiver Operating Characteristic (ROC) curve and the Precision-Recall (PR) curve, which emphasize its ability to accurately detect theft while maintaining a low rate of false alarms. The results demonstrate potential of deep learning techniques in improvising traditional surveillance system, and the approach can be enhanced and adapted for future real-world applications. This method highlights how the deep learning techniques can be used to improve conventional surveillance systems, offers an efficient solution for automated theft detection making it suitable for enhancing safety in different real world environments.