Research On Online Ensemble Learning Method And Its Application In Video Object Detection

Object detection is one of the important research subjects in pattern recognition and computer vision. It has significant application value in many fields, such as military reconnaissance, intelligent transportation, video monitoring, human-computer interaction and so on. In video object detection process, environment interference and object itself variation often appears, and these will result in false detection or leak detection of the detection model and then degrade the detection performance. Based on the depth exploration of the domestic and international research situation of the object detection technique, some researches on object detection method based on online ensemble learning framework were done, and then applied it into video object detection in the complicated environment, and further more, proposed a multi-view face detection method based on online ensemble learning framework. The main works of the dissertation can be organized as follows:1. This dissertation researched on the fundamental theory of online ensemble learning, and made summary of the domestic and international research situation and latest research results. The several online algorithms of classic ensemble learning were introduced in the dissertation, and then the framework and basic implementation steps of online ensemble learning method were introduced in detail, the analysis of its limitations was prepared for follow-up study.2. According to the deficiency of the ensemble classifier designing and automatically online sample labeling in the existing online ensemble learning methods, an adaptive nesting-structed cascade classifier and a sample labeling method based on confidence function were proposed. The nesting-structed cascade algorithm can flexibly adjust the structure in and between the classifier layers, maximize the adaptation to the variation of the object model, accelerate convergence of the classifiers and reduce the number of the weak classifiers in cascade. In the process of the samples’online labeling, the proposed algorithm first proceed preliminary verification combine with a tracking method, and then use confidence function to get the sample class information.3. A multi-view face detection method based on online ensemble learning was proposed to solving the multi-view face detection problem in the complex environment in video. It classified the multi-view face into five categories, such as left all profile face, left half profile face, front face, right half profile face and right all profile face. An initial multi-view face detection model and verification model are trained firstly by a small amount of manually labeled training samples, and then use the detection and verification output to update the detection model based on online ensemble learning method, and update the multi-view face verification model combining with the key frame technique and incremental learning method. The experimental results show that the proposed method can detect multi-view face relatively accurately and it is robust and adaptive under the premise of no obvious decline of the detection speed.