Research Of Object Tracking Based On Distance Metric Learning

Object tracking is an important research branch in computer vision area with multiple ap-plications in our daily life and industrial production. However, there exist various challenges in tracking process, such as illumination variation, occlusion, scale variation, rotation, background clutter and so on, which makes it still a challenging topic to develop a robust tracker.We first introduce the research purpose and summarize the research status. Related distance metric learning algorithms and hashing methods are also introduced, based on which we pro-pose two novel object tracking algorithms, i.e. object tracking algorithm based on collaborative distance metric learning and object tracking algorithm based on collaboratively learnt hashing codes and distance metric.In the first algorithm, strengths of exploiting collaborative appearance models and introduc-ing distance metric learning into object tracking are analyzed. Then the similarity fusion method based cross noise-reduction is proposed based on random walk. We utilize this method to fuse different appearance models and distance metrics and match the candidate samples by the fused similarity. Because the performance of the supervised distance metric learning methods usually rely on the quality of training samples, a preprocess algorithm based on correlation filter is also proposed to select the negative training samples.In the second algorithm, the drawbacks of hashing algorithms are analyzed. To obtain more robust hash codes, we incorporate distance metric learning into spectral hashing method. The hash codes and distance metric are simultaneously learnt in our algorithms. We solve the optimization problem by cross gradient descent. The learnt hash codes are applied in the tracking framework and used as the appearance models when conducting nearest neighborhood search.The two novel object tracking algorithms proposed in this paper achieve favorable perfor-mances against the state-of-the-art algorithms in extensive experiments on various challenging video sequences.