Deep Learning and Sensor Fusion Technique for Indigenous Intelligent Vision System for Theft Detection
M. V. Pranav, Veena N. Hegde, R. Siva Sankara Raju, Roopa S. · 2023
This paper proposes an indigenous system with modified architecture to classify different kinds of objects used in burglary and certain suspicious activities identified in thievery. The classifier used is a part of the security system to detect the images/ movements pertaining to theft or unusual acts while shoplifting valuable objects from jewellery stores. A sequence of integrated real-time computer vision algorithms and sensor fusion technology are implemented. The system adopts the transfer learning method (has 68 layers, 75 connections) as a pre-trained neural network to identify the unusual objects. Here, a person can take suitable action to aid the situation. The classifier algorithm was trained on a dataset of 144 cases of images and augmented appropriately belonging to seven different classes. It provides an accuracy of 97.73% during the validation phase and the confusion matrix resulted in an accuracy close to 95% and an F1-score of 84%.