An End-to-End Dehazing Siamese Region Proposal Network for High Robustness Object Tracking

The haze scenes will bring negative impact to the object tracking process. In haze scenes, aerosols in the air will decrease the information entropy of the image generated by the imaging device. The reduction of information entropy means the loss of image detail information, which will mislead the existing algorithms to extract wrong object feature, causing the failure of tracking. Firstly, this paper explores the problem and proposes an end-to-end Dehazing Siamese Region Proposal Network (DH-SiamRPN). Then, we innovatively design an adaptive transmittance estimation module based on image quality evaluation index, and pre-integrate it into the dark channel dehazing algorithm to preprocess the video sequence. After that, based on the siamese network framework, AlexNet is employed to extract the depth features of the image pair, and the cross-correlation operation of classification and regression is performed according to the regional proposal network (RPN). In addition, inspired by DIoU modeling ideas, we innovatively optimize the proposal selection part. The experimental results demonstrate that the DH-SiamRPN has obvious advantages in precision and speed compared with other state-of-the-art trackers. Its tracking performance meets the requirements of high robustness and speed.