Shape based joint detection and tracking with adaptive multi-motion model and its application in large lump detection.
详细信息
文摘
This thesis is motivated by a practical real application,Large Lump Detection LLD),for which we provide a complete automatic system to detect large lumps in the oil sands mining surveillance videos. To this end,we propose a solution built around three main research components,each of which raises a specific issue,is formulated in a general way,and is tested on both the LLD problem and other similar applications. The first issue is related to the detection of objects that undergo sudden changes in motion. We formulate this problem in a joint detection and tracking JDT) framework using multiple motion models,where these models are predicted adaptively. The prediction exploits the correlation between motion models and object kinematic state. As a result,objects are detected more accurately when they change their motion. The second issue concerns defining an appearance model which differentiates objects from background in an effective manner. We propose a novel shape based appearance model for kernel based trackers which typically model an object with a primitive geometric shape. As a result,by employing the proposed shape based appearance model,the kernel based trackers can improve their accuracy significantly. The last issue aims to ensure an object detection which handles the steam occlusion. We propose a new steam detection method which directly feeds a discrete wavelet transformed image to an Adaboost classifier. In this way,the proposed method is not only accurate because a proper classifier is learned by Adaboost,but also computationally efficient because the feature extraction step is omitted. The complete object detection solution for the LLD problem is obtained by combining the above three techniques. The proposed steam detection method ensures that objects of interest are not occluded,and then,the improved JDT method with the shape based appearance model performs the detection. Extensive experiments and encouraging results which demonstrate the effectiveness of the proposed solution to the large lump detection problem are provided.