DOI: https://doi.org/10.59321/BAUETJ.V4I1.6
AUTHOR(S)
Partha Pratim Debnath1, 2*, Abrar Faiaz Adnan2, Nazmul Hussain2, Md. Abdul Hamid3
ABSTRACT
In this research, the focus is to utilize object detection techniques to track and classify violence in public places in Bangladesh. Based on Google’s TensorFlow, the TensorFlow Object Detection API is an open-source platform that has been utilized to train and evaluate an SSD-MobileNet-v2 model on a dataset specifically created for this purpose, containing images of violent activities in public places from 2000 to 2021. The model is fine-tuned with a training dataset for 4 classes and evaluated on a test dataset. The results show that the performance of detecting armed civilians is the highest with an accuracy of 54.7%. However, the other labels such as setting fire, vandalism and law enforcement were not very accurate and were omitted from the final model to maintain a higher precision for the armed civilian class. The proposed methodology and the obtained results could potentially aid other researchers in designing custom object detectors for similar datasets.