Objects Segmentation And Tracking In Complex Scenarios

In the fields of computer vision, objects segmentation and tracking hasalways been one of the most important topics. With the development ofhigh-performance computers and high definition video cameras, intelligentvideo analysis techniques and more related algorithms are in great demand.Objects segmentation and tracking has close correlation in many applicationfields, including automatic surveillance, video retrieve, human-computerinteraction, navigation, traffic monitoring, and motion based detection andso on. This paper mainly focuses on the topic of objects segmentation andtracking in complex scenarios.The paper consists of three parts. The first part is shape propagatedjoint segmentation with graph matching corroboration for silhouettetracking, which has two main contributions. The first one is that we connectthe shape prediction map with the image frame during the segmentationphase. With the help of relocation coordinate, the novel methods of thegraph construction makes full use of the prior shape information andpropagate such information for latter segmentation, which effectivelyimprove the object segmentation accuracy, and at the same time ensure thetemporal consistency. For the validity of segmentation, our secondcontribution is to design an energy function, and graph matchingtechniques are used to minimize the energy. And it compensates for thesegmentation results if needed and can also cope with occlusion duringtracking.The second part is about pulmonary blood vessels and nodulessegmentation via Vessel Energy Function and Radius-Variable sphere model.It is difficult to classify pulmonary nodules from blood vessels because of the low contrast of the intensity between nodules and blood vessels in CTimages. Therefore, it is a critical problem to detect the nodules that areattached to the pulmonary blood vessel and reduce the FPs of nodulesdetection at vascular. The contribution in this work is that we introduce anovel Vessel Energy Function (VEF), whose energy is relatively highinside a blood vessel while decreases dramatically at vessel boundaries andnodule regions, by which we can accurately segment the blood vesselregion. Pixels with high VEF will be set as blood vessel regions, andpropagation will cease at vessel boundaries where VEF is relatively low, sothat we can obtain the whole blood vessel region. A radius-variable spheremodel is also utilized to refine the segmentation results and ensure thesmoothness and continuity of the blood vessel centerline.The last part of this paper is objects detection and tracking based onfeature points matching. A real-time detection and tracking system isproposed in this work. During the detection phase, we introduce adetection algorithm combining Harris corner and SIFT descriptor.Adaptive scale factor and adaptive threshold for corner detection areproposed to reduce the time complexity and improve the detectionefficiency. And in the tracking phase, we use LK optical flow to track thestable feature points. The feature extraction algorithm is improved to fit thetracking scheme. Feature points are re-extracted when occlusion happens,so as to deal with partial or fully occlusion. The experimental results showthat the proposed detection and tracking system can reach the real-timeperformance on Windows, and near real-time on Android OS.