Object Tracking Using On-line Distance Metric Learning
D. Cheng, J. Hu, W. Lyu
In order to improve the real-time quality and precision of object tracking, an algorithm using distance metric learning is studied. First, instances are selected around the objects, and features vectors are extracted by using the compress sensing theory. Second, distance metric is trained according to the random projection theory. Finally, the Mahalanobis distance of target object and possible instance are calculated, and the instance which has the smallest distance from object is the tracking object. Experiments on three videos show that the calculating load of the compressed features is ¾, which is less than the one using the uncompressed ones. Using the trained distance metric to calculate the location of target improves the tracking precision.