Algorithm On Multi-Camera Multi-Object Tracking In Complex Environment

Tracking of objects is important for many computer vision applications. Currently, most of algorithms in object tracking are applicable to single-camera single-object tracking. Those algorithms will failed when there are multiple objects in complex and crowed environments. For example, mean shift algorithm, particle filter, Kaman filter and optical flow method are all fit for single object tracking, when tracking multiple objects, the computational cost is large, and can’t meet the real-time application requirements.This paper presents a multi-feature multi-level data association approach to robustly track multiple objects in complex and crowed environments. There are three main advantages of this approach: first, many features are integrate into the affinity model to measure similarity between detection responses or between tracklets in the association optimization framework; second, associations are made in several levels and the affinity measure is refined at each level based on the knowledge obtained at the previous level; Third, occlusion by other objects or scene structures can be detected, and adopt different methods to resolve occlusion according to the size of occlusion. Experiments are carried out by tracking pedestrians in challenging datasets. Compared to previous methods, the results show a great improvement in both tracking accuracy and speed.Video sensor networks have attracted much interest in recent years due to the broad coverage of an environment and the possibility of coordination among different cameras. Although the field-of-view (FOV) of a single camera is limited and cameras may have overlapping or non-overlapping FOVs, seamless tracking of moving objects can be achieved by exploiting the hand-off capability of multiple cameras. In order to resolve occlusion problems, multiple cameras were used to obtain continuous visual information of people in order to track through interactions. The algorithm is based on the theory of binocular vision, by calculating the fundamental matrix F, we associate object through multi-cameras. At the meantime, a camera assignment and hand-off model on a set of criteria to choose the“best”camera for each object in the scene were proposed.Multi-camera multi-object tracking when the overlapping FOVs between cameras are small were also researched. An algorithm was adopted to eliminate the impact of illumination difference in multiple cameras that using multiple cues to recover shading and reflectance intrinsic images from a single image. After that, the methods based on the objects’location and camera’s horizon line were used to implement multi-object tracking among multi-cameras. Results demonstrated that the system can maintain people’s identities by using multiple cameras cooperatively.