New Thermal Automotive Dataset for Object Detection

Abstract

Although there are many efficient deep learning methods, object detection and classification in visible spectrum have many limitations especially in case of poor light conditions. To fill this gap, we created a novel thermal video database containing few thousands of frames with annotated objects acquired in far infrared thermal spectrum. Thanks to this we were able to show its usability in the traffic object recognition based on the YOLOv5 network, properly trained to gain maximal performance on thermal images, which contain many small objects and are characteristic of different properties than the visible spectrum counterparts. The proposed thermal database, as well as the fully trained model are main contributions of this paper. These are made available free for other researchers. Additionally, based on the highly efficient car detector we show its application in the car speed measurement based exclusively on thermal images. The proposed system can be also used in the Advanced Driver-Assistance Systems (ADAS), and help autonomous driving.

Real-Time Detection of Small Objects in Automotive Thermal Images with Modern Deep Neural Architectures

“…In a previous article [4], a thermal automotive dataset was introduced, specifically designed for object detection using the YOLOv5 model [5]. However, the presented dataset had certain limitations, as it contained only images captured during winter conditions.…”