Event-based object tracking framework robust to motion blur and severe illumination

Abstract

Event-based single object tracking framework robust to motion blur and severe illumination is proposed in this paper. Conventional single object tracking researches use RGB images and achieved high tracking performance with the aid of deep learning. However, the tracking performance significantly drops when the input frame is not clear due to the afterimage or extreme lighting. In this paper, the event camera that captures per-pixel brightness changes with high speed is used to handle those challenging situations. In the training stage, the list of events is aggregated with the learnable kernel and the tracking network is trained via similarity learning. In the testing stage, the target is tracked with the pretrained network and static module. The evaluation results on the generated synthetic event dataset with the event simulator and the captured real event dataset demonstrated that the proposed method robustly track the targets in the scene with motion blur and severe illumination.