Publication

Abstract : Identification of objects in an image plays a crucial role in UAVs, self-autonomous vehicles, and other applications. It is vital to have the capability to navigate and identify objects at any time of the day, especially at night, when we encounter situations of low-light and dark conditions. This necessitates the ability to have a proper object detection model to predict the object in these conditions. We introduce a way to detect and classify objects in images taken with thermal cameras using different image processing techniques, pattern recognition, and machine learning algorithms. We have done a comparative study of two different models with appropriate hyper-parameter tuning that allows us to choose the right model for their application based on the constraints involved in the problem. We present an extensive study of the models based on the COCO evaluation metrics and other important loss metrics. Experimental results on the FLIR dataset confirm that we can use the pre-trained models in object detection to train, identify, and label objects in thermal images.