A semiautomatic method for pedestrian ground truth generation in thermal infrared on-board videos

Currently, pedestrian detection and tracking algorithms of Thermal Infrared (TIR) on-board videos encounter lack of comprehensive pedestrian datasets for benchmarking. The generation of ground truth is a tedious and error-prone task in the process of creating the dataset of annotated videos. This paper puts forward a novel semiautomatic video annotation method to facilitate annotating pedestrians in TIR on-board videos. The proposed method consists of two phases. In the first phase we learn the pedestrian appearance models online, then in the second phase we use the learned models to automatically annotate the pedestrian in the other frames. We present a video annotation tool to verify the effectiveness and reliability of our method. A comparison between our tool and the state of the art of onboard video annotating tools was performed, which showed how our annotation tool provides a high ground truth quality with shorter annotation time when annotating pedestrians in TIR on-board videos with bounding boxes.