Shadow-aware Monochrome Video Object Tracking

Video object tracking is to extract features of the object through a series of visual methods and automatically estimate the position and size of the object in subsequent frames,given only the location in the initial frame.In recent years,siamese trackers have achieved a good balance in accuracy and speed,and have become the mainstream research direction in object tracking field.The existing siamese trackers mainly study the tracking of any object in any common scenario.However,in actual scenarios,due to the existence of lighting,the tracking accuracy is often affected by the shadows generated by the objects.In addition,these algorithms are usually designed for color videos.But for some specific task scenarios,it is necessary to track the object in monochrome videos.For monochrome videos,object tracking is more difficult due to low contrast and missing color information.On the basis of summarizing and analyzing the research status in object tracking field,this thesis focuses on the problem of object tracking in monochrome videos containing shadows.The major works of this thesis are summarized as follows.(1)A shadow-aware monochrome video object tracking algorithm is proposed to reduce the influence of shadows on the tracking accuracy of monochrome videos.The algorithm network framework is generally composed of a siamese network for feature extraction and a region proposal network for proposal generation.The classification and regression results of proposals are used to determine the position of the object.(2)A shadow-aware module based on shadow prior is proposed in the siamese subnetwork.This module can apply the features extracted from the shadow prior branch to the features extracted from the image branch in a spatial attention manner.It enhances the features comparison between shadow and non-shadow regions in the image branch.(3)A selective feature fusion module is constructed to extract more fine features.This module can adaptively adjust the size of the receptive field according to different objects.It obtains more discriminative features for the object,thereby further improving the performance of the tracker.(4)Aiming at the particularity of monochrome videos,a data augmentation method that randomly changes the brightness or contrast of the detection image is used in the training data.In this way,the proposed model learns more robust feature representations for illumination changes in monochrome videos,while improving the tracker's anti-interference ability to object shadow areas.(5)In addition,there are fewer video sequences containing shadows in the public object tracking datasets,so a new dataset SSOT(Shadow Scene Object Tracking)is constructed to better verify the effectiveness.The dataset is obtained by shooting videos in the scenes with shadows and labeling true object boxes.It mainly consists of 53 video sequences,including scenes of buildings,trees,racers and so on.The proposed algorithm is implemented by using Py Torch deep learning framework under Ubuntu 16.04 system environment.The proposed model is tested on monochromatic public object tracking video sequences and the newly constructed SSOT dataset,and analyses are performed by being compared with existing object tracking algorithms.The experimental results show that the proposed algorithm not only achieves excellent tracking performance in monochrome video sequences with shadows,but also can track the objects well in monochrome video sequences without shadows,which efficiently verifies the effectiveness of the proposed algorithm.