A Study on Lightweight Object Detection in Thermal Images and Its Recent Advances

Object detection is a vast domain which comprises of various methods of detecting objects efficiently and with good accuracy. Most of the images that we use for detection of objects are RGB images, but what if the images that we are using are captured during adverse weather conditions like foggy, hazy or in bad illumination conditions (during night)? Our model for object detection might not perform well in such cases. To overcome the problem of illumination, images can be captured by an infrared (IR) camera which are robust under such conditions. Thermal images are captured through infrared cameras and they capture the image with the radiation (heat) into visible images, such images are useful in many applications like autonomous driving, military applications, object detection in adverse conditions, etc. It is essential to make use of thermal images and exploit its use to solve many real-world problems faced in object detection. This review tries to analyze the various existing models and design a lightweight network framework for detecting objects under adverse weather conditions in thermal images. The work will primarily focus on comparing various models and review it based on efficiency and appropriateness for specific use cases in thermal dataset. The applications of the robust deep learning models like you only look once (YOLO), convolution neural network (CNN), etc., are analyzed through experimentations and based on the results we find the shortcomings in thermal image object detection space and try to fill the gaps by suggesting a framework of methodology to detect an object and further compress the model using techniques like pruning and knowledge distillation to make the model lightweight so that the research can be extended in this domain even without compromising on the accuracy and disk space to do object detection in thermal imagery. This research can be utilized in various fields to improve accuracy in detection of objects.