基于多核多示例学习的洗车行为识别方法研究

发布时间：2018-09-12 06:59

【摘要】：洗车行为识别是复杂场景下人体行为识别的一个分支,目前简单场景下的人体简单动作的识别已基本得到解决,而复杂场景下的行为识别仍面临很多困难。洗车行中摄像头的特殊角度拍摄人体动作轮廓不分明以及汽车、工人频繁移动造成的“鬼影区域”都使得对洗车行为的识别更加困难。当前传统的行为识别算法并不能适应洗车行这种特殊环境,针对洗车行为识别,本文提出了一种基于多核多示例的学习算法,以提高洗车行环境下洗车工人行为识别的准确率。本文采用了改进的ViBe背景差分法对运动目标实时检测来解决消除“鬼影区域”问题,采用HOG-LBP特征提取算法来应对人体动作轮廓不分明的问题。识别算法采用多核多示例学习算法,该算法将多核支持向量机与多示例学习算法有机结合,能够有效的处理提取到的HOG-LBP融合特征,同时也提高了识别算法的学习能力,进一步提高洗车行为识别的准确率。实验证明,多核多示例学习算法在实验构建的洗车行为数据集上,与传统的行为识别算法相比,其识别率有所提高。本文中的算法是针对洗车行环境提出的,也同样适用于类似洗车行环境的复杂场景下的行为识别问题。
[Abstract]:Car washing behavior recognition is a branch of human behavior recognition in complex scenes. At present, the recognition of simple human actions in simple scenes has been basically solved, but the behavior recognition in complex scenes still faces many difficulties. The special angle of the camera in the car washing shop makes it more difficult to recognize the car washing behavior because of the unclear outline of human body movement and the "ghost area" caused by the frequent movement of the workers. At present, the traditional behavior recognition algorithm can not adapt to the special environment of car washing line. In view of car washing line recognition, this paper proposes a learning algorithm based on multi-core and multi-example to improve the accuracy of car washing workers' behavior recognition in car washing environment. In this paper, the improved ViBe background differential method is used to detect moving objects in real time to solve the problem of eliminating the "ghost region", and the HOG-LBP feature extraction algorithm is used to deal with the problem of unclear human action contour. The recognition algorithm adopts multi-core and multi-example learning algorithm, which combines multi-core support vector machine with multi-example learning algorithm, which can deal with the extracted HOG-LBP fusion features effectively and improve the learning ability of the recognition algorithm. Further improve the car washing line for recognition accuracy. The experimental results show that the multi-core multi-example learning algorithm is more efficient than the traditional behavior recognition algorithm in the experimental data set. In this paper, the algorithm is proposed for car washing environment, and it is also applicable to the problem of behavior recognition in complex scenarios similar to car washing environment.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U472.2;TP391.41

【参考文献】