Next-Generation Crime Detection and Transmitting: Evaluating Pre-Trained CNN Models
V Viswanatha, A C Ramachandra, Bidare Divakarachari Parameshachari, Sai Manoj Thota, Satya Dev Nalluri, Aishwarya Thota · 2024
Violence detection has garnered significant attention as there’s a growing demand for automated methods to identify violent actions. This surge in interest stems from the utilization of surveillance cameras in diverse locations, where visual data is analysed to detect instances of violence. The project’s objectives are tailored to address key research gaps within the realm of crime detection. Firstly, it aims to refine a pre-trained model through meticulous fine-tuning, culminating in the development of a robust crime detection algorithm. Secondly, the project endeavours to transcend theoretical realms by deploying this model in real-time scenarios, seamlessly integrating it with live camera feeds to enable continuous monitoring and analysis. Lastly, the project will conduct a thorough comparative analysis of various models, discerning their strengths and weaknesses to pinpoint the optimal solution for effective crime detection. Through these objectives, the project aspires to contribute significantly to the advancement of crime detection technologies. The methodology employed includes collecting images for the dataset, cleaning and pre-processing the data, augmenting it for diversity and it increases dataset diversity through transformations. Implementation is done of transfer learning with pre-trained models like ResNet or VGG, and training the model to learn patterns and relationships within the data for accurate predictions or classifications. Model training is essential for the model to learn patterns and relationships within the data, minimizing prediction errors. The results reveal a peak accuracy F1-score of 95% achieved with ResNet. A high F1score of 96% indicates strong performance in accurately identifying positive instances while minimizing false positives and false negatives.