Moving Objects Tracking In Aerial Videos

Visual tracking is a hot topic computer vision that received a lot of attention during the last decade.Most of the tracking approaches use data association methods based on an online tracking-by-detection strategy to obtain the trajectories of tracked objects.It is still a difficult problem due to illumination changes,object's shape/size changes,occlusion,and the complexity of the background.In addition to these,multi-object tracking methods must also be able to deal with data association(correspondences between two frames),changing numbers of objects,inserting new objects and deleting them when appropriate.In this thesis,we investigate robust long-term tracking algorithms to solve unreliable objects detections and occlusions in airborne videos,with the aim of building an automatic system that can accurately localize moving objects that appear in aerial video frames and continuously track them.This thesis puts emphasis on moving objects detection and tracking in different extreme scenarios such as occlusion due to platform viewpoints and static scene objects,and changes in the size and appearance of the targets.It breaks down into three parts.1.A Dynamic Weighted Compressive Model(DWCM)was proposed for object tracking,for which we based our development on the recently proposed Compressive Tracking(CT)method.CT is a robust tracking algorithm for which the object's appearance model is based on non-adaptive random projections and the tracking problem is treated as a binary classification between target and background.However,CT trackers cannot deal well with serious occlusions,illumination variation and the object's scale changes due to its constant tracking box.To solve these shortcomings,our proposed model adopts multiply sparse measurement matrices along with weight map to preserve important features to build the object appearance model.The weight map combines a contrast weight and a feature weight is used to select the best measurement matrix that preserves the image structure of the targets during tracking.The experiment results demonstrate the compressive sensing with an effectively feature selection is efficiently to reduce the tracking drifting risk in airborne videos.2.An improved compressive tracking based on multi-scale feature selection and feedback strategy with the Pixel-wise Learner(PLCT)was proposed for object tracking.To reduce the complexity and computation of multiply matrix projections in DWCM as well as to increase the robustness of the CT-based tracker under scale variation,background clutter and deformations in airborne videos,a Pixelwise classification stage is incorporated in the PLCT to obtain a relatively stable appearance model,by distinguishing object pixels from the background.In addition,we identify potential distracting regions that are used in a feedback strategy to handle occlusion and avoid drifting toward nearby regions with similar appearances.We evaluate our approach on several benchmark datasets to demonstrate its effectiveness with respect to stateof-the-art tracking algorithms.3.A four-stage Hierarchical Association framework was proposed for multiple object Tracking in Airborne video(HATA).In this framework,the detections are associated to tracklets to build the final object trajectory.The tracklets and detections are divided into several groups based on cues obtained from the association results provided by tracklet confidence.In each association stage,different sets of tracklets and detections are associated to handle the following problems: local tracklet generation,local trajectory construction,global drifting tracklet correction and global fragmented tracklet linking.The proposed framework combines data association-based tracking and the proposed PLCT-based tracking,along with object reidentification.Qualitative and quantitative comparisons with state-of-the-art methods,using challenging airborne video datasets,demonstrate its performance.Experiments with challenging airborne video datasets show significant tracking improvement compared to existing stateof-the-art methods.