Integration of texture and depth information for robust object tracking

We proposed a robust object tracking invariant object appearance variations and background clutter. Not only texture information is used, the depth information, which is important to classify the object from complicated background, is integrated as well. For the texture information, multiple instance learning with boosting algorithm is applied to select discriminant Haar-like features between the positive (object) and negative (background). Via the Bayes rule, the posterior probability can be estimated for each sample candidate at each frame. Additionally, we propose two depth-based models that can compensate each other to tackle not only appearance variants but the drift problem as well. Since the depth information is used, the tracking problem is transferred from 2D flat images to 3D stereoscopic image and the tracking results can be more accurate. Because improper model update will cause incorrect tracking results in the following frames, especially the textureless depth templates increase this risk; one specific update mechanism for depth model is proposed to avoid incorrect model update. In the experiments, we design six challenging video sequences to validate the robustness of our system and meanwhile compare our system with the existing famous algorithms. Not only qualitative results but quantitative results are provided for performance analysis. From the experimental results, the proposed system can provide more reliable performance to be invariant to most appearance changes and has more accurate tracking results than other famous algorithms.