Vehicle detection in thermal images with an improved yolov3-tiny

In driver assistance systems, thermal cameras are often used because they can provide compensation information for other sensors in the case of darkness or glare. For many existing thermal image vehicle detection algorithms, they can get a good detection accuracy in some occasions but the detection speed is relatively slow so they can't meet the real-time requirements. In order to address this problem, we proposed a real-time thermal image vehicle detection algorithm based on yolov3-tiny. We have made two major improvements to the original yolov3-tiny. The first is to recalculate the anchor box priors by running k-means clustering algorithm on the bounding boxes of training dataset to make the network easier to learn. The second is to deepen the network structure of the original yolov3-tiny, so that it can better extract the characteristics of the vehicle in thermal images, so as to improve the vehicle detection accuracy. Our experimental results show that the mean Average Precision (mAP) of our proposed method is 6% higher than the origin yolov3-tiny while maintaining the detection speed comparable to the original yolov3-tiny.