基于多特征融合的视频足球比赛中的团队行为识别方法研究

项目名称： 基于多特征融合的视频足球比赛中的团队行为识别方法研究

项目编号： No.61462008

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 王智文

作者单位： 广西科技大学

项目金额： 48万元

中文摘要： 视频足球比赛中的团队行为识别易受遮挡、光照、噪声、视角以及多尺度等因素的影响，且要在复杂场景下实时鲁棒地识别团队行为，挑战极大。针对这些挑战，本项目研究用自动生成径向基函数神经网络来融合提取的团队行为的多种特征，解决因提取的特征过多导致计算的复杂性增加。研究用权重化条件随机场模型为视频足球比赛中的团队行为进行建模，用代数的线性空间取代非线性流形空间来大大简化建模过程中的计算量。拟采用多尺度自适应选择的特征点检测方法来克服视频足球比赛中的光照和尺度变化的影响。设计基于迁移学习算法的局部时空码本原型构建算法，解决遮挡和多视角问题。设计出团队行为识别的模糊推理系统实时鲁棒地进行视频足球比赛中的团队行为识别，解决在复杂场景中同一场合中的多种行为识别的问题。研究使用基于先验知识和人工神经网络的树结构混合分类器对团队行为进行分类，提高分类的准确性。本项目的研究具有极大的经济价值和社会价值。

中文关键词： 团队行为识别；特征提取及融合；迁移学习算法；权重化条件随机场模型；模糊推理系统

英文摘要： It is a great challenge for recognizing the team actions robustly and in real time and in complex scenes, because team actions recognition for video football game is vulnerable to occlusions, light, noise, perspective, multi-scale and other factors. In response to these challenges, automatic generation radial basis function neural network is researched to fuse extracted features of team behaviors to address excessive computation complexity caused by feature extraction in this project. Weighted CRFs model is researched to model for team actions of video football game and the algebra linear space is used to replace the nonlinear manifold space to simplify the calculation of modeling. Feature point detection method using multi-scale adaptive selection is intended to use to overcome the effects of the change of illumination and scale in video football game. The constructed algorithm of the prototype of local space-time code based on transfer learning algorithm is designed to solve the problem of occlusion and multi-angle. The fuzzy inference system for team actions recognition is designed to identify robustly and real-time team behaviors in video football game, solving the problems of identifying multiple behaviors occurring on the same occasion in complex scenes. Artificial neural network hybrid classifier based on prior knowledge and tree structure is designed to classify team behaviors and improve the accuracy of classification. Research of this project has great economic value and social value.

英文关键词： Recognition of team actions;Feature extraction and integration;Transfer learning algorithm;Weighted conditional random fields model;Fuzzy inference system