A Robust Vehicle Detection Method in Thermal Images Based on Deep Learning

Thermal cameras have been widely used in driver assistant system to provide supplementary visual information in darkness. As a consequence, vehicle detection in thermal images is highly required. Current methods for thermal image vehicle detection can hardly make a good balance between accuracy, speed and robustness. In this paper, we propose a deep-learning-based method for automatic vehicle detection in thermal images which employs yolov3-tiny as the main backbone with three major improvements. Firstly, an additional detection layer is designed and added to the original yolov3-tiny network for small vehicle detection. Secondly, a group of side-way blocks are introduced to extract and combine features in different layers so as to improve the overall detection accuracy. In the end, the Spatial Pyramid Pooling (SPP) module is adopted for multi-scale feature integration to further improve the detection performance. Experimental results show that the proposed method raises the mean Average Precision (mAP) by 4% compared to the current state-of-the-art method and 10% compared to the original yolov3-tiny network with little increase in time-consumption. In conclusion, the evaluation results suggest that our method is robust and effective for vehicle detection in thermal images.