The Recognition Of The Fall Action Indoor

The traditional video surveillance system usually records the scene for the investigationand proof submission afterwards, which is not only a lot of work, costs huge human andmaterial resources, but also has low precision. Therefore, it is meaningful to detect andanalyze the abnormal epic and object in the scene with the video automatic analysistechnology, among them the motion analysis based on video surveillance becomes an activeresearch field in the computer vision nowadays. During the recent years, with the speed-up ofthe aging process in China, the proportion of seniors quickly climbs, while “falling” is oneimportant factor that leads to death, it is particularly true that “falling” leads to death for theseniors who live alone. Therefore, we use the video automatic analysis technology to detectand recognize the indoor fall action in this project.This paper focuses on many aspects of the motion analysis system based on videosurveillance, mainly includes the construction of the video data library, the presentation ofmotion, fall action detection and detect fall action in the consecutive video. We take themodularity design method and divide the system into motion target extraction, motion featureextraction and fall action detection. On the basis of the constructed video dataset, we do someresearch on the fall action and the three common family actions such as squat, bend and sit,which are easily confused to fall, Then we use the machine learning method to effectivelydistinguish fall from squat, bend and sit. Furthermore, we make full use of the heightinformation to orient fall, squat, bend and sit in the consecutive video, and then use themachine learning method to detect fall action.The work of this paper is as follows.1. Record video data, do the motion target extraction: When we use video automatictechnique to recognize fall action, squat, bend and sit easily confuse to fall. Therefore, werecord the video of fall, squat, bend and sit. Here we regard the motion process from the mostheight to the least height as our research target. Then we segment the video, and use the framedifference method to extract the target, afterwards, we take the morphology and district areaways to optimize the result. Finally, the video dataset is constructed. 2. Feature extraction module: We first do some specific researches on the four confusedactions, and then decide to use the feature extraction with machine learning method to detectfall action. We extract the normalized height and normalized height and width feature, blockgradient direction feature, Huinvariant moment feature. As for the normalized height andnormalized height and width feature, we use the motion level, motion gradual change tooptimize the first feature. Besides, we do some statistical analysis for the second and thirdfeatures and demonstrate the stability within one certain action and the difference among thefour different actions.3.Fall action detect: We use the machine learning method to detect fall action. Wechoose SVM to detect fall action on the basis of the small sample characteristic of the fouractions. We learn the method and theory about SVM in detail, and then learn the principleabout the kernel function, grid-searching method for parameter choice and the experimentmethod of cross validation. Then we do the cross validation experiment for the562segmentedsamples, the overall recognition rate of the fall action is up to93.25%.4. The process towards successive video: We combine the frame difference and gradientmethod to accurately get the height of the target, and then orient the starting and finishingpoints of the four target actions such as fall, squat, bend and sit with the height information,finally we do the feature extraction and fall action detection.As to the detect of the fall action indoor, On the basis of the constructed video dataset,we do some research on the fall action and the three common family actions such as squat,bend and sit, which are easily confused to fall, Then we use the machine learning method toeffectively distinguish fall from squat, bend and sit. Furthermore, we make full use of theheight information to orient fall, squat, bend and sit in the consecutive video, and theneffectively distinguish fall from squat, bend and sit in the consecutive video.