Fusion of Thermal and RGB Images for Traffic Object Detection

Conventional approaches to traffic data collection often involve the use of RGB cameras. In an outdoor scenario, the data collected is affected by many factors, such as poor/excessive lighting, weather, occlusion, etc. Introducing a thermal sensor helps tackle the loss of information in RGB images. This paper introduces a novel algorithm, the 4-channel YOLO V-8 model, to combine the strengths of RGB and thermal images into a single 4-channel input for object detection. This paper develops and implements the 4-channel model and compares it with some of the existing sensor fusion approaches, such as pixel-level, feature-level, and decision-level fusion. These algorithms are evaluated on the FLIR ADAS dataset. Results show the superior performance and reduced loss in information of the new 4-channel algorithm over the existing algorithms.