基于视觉的运动人体特征描述与行为识别研究

发布时间：2018-05-01 19:33

本文选题：计算机视觉 + 目标检测　；参考：《浙江工业大学》2015年博士论文

【摘要】：基于视觉的运动人体行为识别是计算机视觉和机器学习等领域的热门研究方向,在智能监控、人机交互等方面有着十分重要的应用价值。本文在运动目标的检测、特征提取和描述、人体行为识别三个方面进行了研究,主要工作和成果如下:1.在运动目标检测上,本文以视觉背景提取算法(Vi Be)算法为基础,提出一种鲁棒的自适应视觉背景提取算法(RAViBe)。RAViBe采用鲁棒的基于图像对比度的背景建模方法,采用自适应的背景更新与目标分类方法来进行运动目标检测。实验结果表明,与ViBe算法相比,RAVi Be算法在保证了算法简洁高效的同时,在运动目标检测中能获得更好的检测效果,能快速有效地消除鬼影区域,并且在光照变化的情况下保证算法的鲁棒性。2.在特征提取和描述上,本文提出一种基于块直方图的融合特征描述子(BBHFFD),提取运动区域中的剪影特征与光流特征合并成融合特征,并采用基于块直方图的方式加以描述。在此基础上,本文采用判别公共向量(DCV)的方法对融合特征进一步降维,并提出一种改进的快速判别公共向量(FDCV)的方法,用标量分类代替原算法中的向量分类。实验结果表明,与单一特征相比,采用融合特征在行为识别时能获得更高的识别率;同时,与DCV相比,采用FDCV能在保证识别率的前提下,将目标分类的计算速度提高约两倍。3.在基于状态空间的人体行为识别上,本文将SVM-HMM混合模型引入人体行为识别中,用属于判别式模型的支持向量机(SVM)来替代属于产生式模型的GMM模块,使得在训练样本不足的情况下能获得更优的分类性能。实验结果表明,与GMM-HMM相比,SVM-HMM混合模型在短序列视频中获取了更高的识别率。而在长视频序列的识别上,得益于HMM对时间序列的建模能力,SVM-HMM混合模型同样取得了较好的识别率。4.在基于规则的人体行为识别上,本文提出一种基于遗传算法规则进化的人体行为识别方法,为人体行为设定一组约束条件作为规则,同时将遗传算法应用于规则的更新上,使规则能得到不断地进化从而获得自适应能力。实验结果表明,与初始规则相比,进化得到的规则能在人体行为识别中获得更高的识别率。5.结合以上的研究工作,针对水电站异常行为监控的应用环境,在windows平台下结合OpenCV实现了人体异常行为监控系统。系统通过训练,能在固定场景中对异常行为进行识别并显示报警信息。
[Abstract]:Motion human behavior recognition based on vision is a hot research direction in the fields of computer vision and machine learning. It has very important application value in intelligent monitoring and human-computer interaction. In this paper, the detection of moving targets, feature extraction and description, and human behavior recognition are studied. The main work and results are as follows: 1. In moving target detection, based on the visual background extraction algorithm (Vi Bee), a robust adaptive visual background extraction algorithm, RAViBe.RAViBe, is proposed, which adopts robust background modeling method based on image contrast. Adaptive background updating and target classification are used to detect moving targets. The experimental results show that compared with the ViBe algorithm, the ravi be algorithm not only ensures the simplicity and efficiency of the algorithm, but also can obtain better detection effect in moving target detection, and can eliminate the ghost region quickly and effectively. Moreover, the robustness of the algorithm is guaranteed under the condition of illumination change. In terms of feature extraction and description, this paper presents a fusion feature descriptor based on block histogram, which extracts silhouette feature and optical flow feature from moving region and combines them into fusion feature, which is described by block histogram. On this basis, this paper uses the method of discriminating the common vector (DCV) to further reduce the dimension of the fusion feature, and proposes an improved method of fast discriminating the common vector / FDCV, which uses scalar classification instead of the vector classification in the original algorithm. The experimental results show that compared with single feature, the fusion feature can obtain higher recognition rate in behavior recognition, and compared with DCV, using FDCV can improve the speed of target classification by about twice as much as that of DCV. In the aspect of human behavior recognition based on state space, this paper introduces SVM-HMM hybrid model into human behavior recognition, and uses support vector machine (SVM), which belongs to discriminant model, to replace the GMM module which belongs to the production model. The better classification performance can be obtained under the condition of insufficient training samples. The experimental results show that the hybrid SVM-HMM model achieves a higher recognition rate in short sequence video than GMM-HMM. In the recognition of long video sequences, the good recognition rate of SVM-HMM model is also obtained because of HMM's ability to model time series. For human behavior recognition based on rules, this paper proposes a human behavior recognition method based on evolution of genetic algorithm rules, which sets a set of constraints for human behavior as rules, and applies genetic algorithm to update rules. So that rules can be constantly evolving to achieve adaptive capabilities. The experimental results show that, compared with the initial rule, the evolutionary rule can obtain a higher recognition rate of .5. in human behavior recognition. Combined with the above research work, the human abnormal behavior monitoring system is implemented under the windows platform combined with OpenCV in view of the application environment of abnormal behavior monitoring in hydropower station. Through training, the system can identify abnormal behavior and display alarm information in fixed scene.
【学位授予单位】：浙江工业大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】