人的摔倒动作检测方法的研究

发布时间：2018-04-16 10:39

本文选题：摔倒检测 + 滑动窗口　；参考：《辽宁大学》2017年硕士论文

【摘要】：我国经济社会不断发展,人口老龄化愈加严重,对自动化摔倒检测有着迫切的需求。摔倒检测方法主要可以分为两个流派:第一种是基于物理传感器件的可穿戴式设备检测方法;第二种是近年来新兴的基于视频图像处理的方法。可穿戴设备虽然可以精确的获得人体相应关节的运动参数,但由于其产生了穿戴负担,所以对于老年人并不是十分适用。图像处理和模式识别近年来迅猛发展,将视频处理技术应用于摔倒检测的条件已经成熟。基于视频的检测方法由于其对受众的方便性,在未来会有广阔的应用前景。目前很多基于机器视觉的检测方法都遵循先定位后追踪的方式来获取人体运动情况,这种方式面临四个突出的问题:1、定位跟踪运算量庞大;2、只能通过获得人体轮廓信息来推断人体状态;3、定位视频中摔倒起始位置困难;4、跟踪过程容易跟错目标。为了解决上述四个问题,本文采用了基于滑动窗口的视频事件表示方法,该方法的突出优点是将视频内容抽象成时空节点集合,由于时空节点是对视频内容的完整抽象,一方面可以将原先在原视频上的定位跟踪替换成对时空节点的搜索;另一方面可以做到基于内容的搜索,而非通过人体轮廓信息推断人体状态。针对时空路径搜索问题,我们将原先应用于视频事件检测的时空路径搜索算法在摔倒检测上进行了拓展应用,利用该算法可以在局部迭代过程中求全局最优解和低时间复杂度的优点解决了在未知起点情况下进行快速时空路径搜索问题。对于在复杂场景检测中出现的路径偏差问题,经研究发现主要是由于帧间节点在连接过程中缺乏预测指导所致。根据摔倒动作的特性,我们将摔倒动作分阶段拆解,用马尔科夫过程模型来表示整个动作过程,进而提出了能够增强时空路径搜索中帧间节点局部连接性的转移概率改进方法。在帧间节点的连接过程中依据转移概率对下一帧节点的选择提供指导,解决了路径搜索过程中帧间节点关联性不强的问题。为了对算法的有效性进行验证,我们制作了新的多场景摔倒数据集,并将上述算法在数据集上进行了验证。实验结果表明本文提出的改进方法可以有效的应用在数据集上,并且能获得比较好的检测效果。
[Abstract]:With the development of economy and society, the aging of population is becoming more and more serious.Fall detection methods can be divided into two main schools: the first is a wearable device detection method based on physical sensing devices; the second is a new method based on video image processing in recent years.Although wearable devices can accurately obtain the motion parameters of the corresponding joints of the human body, they are not very suitable for the elderly because of their wearable burden.With the rapid development of image processing and pattern recognition in recent years, the condition of applying video processing to fall detection is ripe.Because of its convenience to the audience, the video detection method will have a broad application prospect in the future.At present, many detection methods based on machine vision follow the method of locating first and then tracking to obtain human motion.In this way, there are four outstanding problems: 1, the operation of location tracking is huge, the human body status can only be inferred by obtaining human contour information, the position of falling down in the location video is difficult and the tracking process is easy to follow the wrong target.In order to solve the above four problems, this paper adopts the method of video event representation based on sliding window. The outstanding advantage of this method is that the video content is abstracted into a set of space-time nodes, because the spatio-temporal node is a complete abstraction of the video content.On the one hand, the original location tracking on the original video can be replaced with the search of the space-time node; on the other hand, the content-based search can be achieved, instead of inferring the human body state from the human contour information.In order to solve the problem of spatio-temporal path search, we extend the application of space-time path search algorithm used in video event detection in fall detection.The algorithm can solve the problem of fast spatio-temporal path search with unknown starting point by using the advantages of global optimal solution and low time complexity in the local iterative process.For the problem of path deviation in complex scene detection, it is found that it is mainly due to the lack of prediction guidance in the connection of inter-frame nodes.According to the characteristics of falling motion, we divide the falling action into stages and use Markov process model to represent the whole action process, and then we propose an improved method which can enhance the local connectivity of inter-frame nodes in space-time path search.The selection of the next frame nodes is guided by the transition probability in the connection process of inter-frame nodes, which solves the problem that the inter-frame nodes are not correlated strongly in the course of path search.In order to verify the validity of the algorithm, we made a new multi-scene fall dataset and verified the algorithm on the dataset.The experimental results show that the improved method proposed in this paper can be effectively applied to the data set and can obtain a better detection effect.
【学位授予单位】：辽宁大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP212.9

【参考文献】