7331

6837

كارشناسي ارشد

معماري كامپيوتر

اصفهان: دانشگاه صنعتي اصفهان، دانشكده برق و كامپيوتر

1391

دوازده، 99ص.: مصور، جدول، نمودار

ص.ع. به فارسي و انگليسي

شادرخ سماوي

6/9/91

بهزاد نظري، محمدرضا احمدزاده

1396/09/21

كتابنامه

برق و كامپيوتر

مهندسي برق و كامپيوتر

ID6837

به فارسي و انگليسي: قابل رويت در نسخه ديجيتالي

Automatic Human Fall Detection System in Video Sequence Behzad Mirmahboub b mirmahboub@ec iut ac ir September 19 2012 Department of Electrical and Computer Engineering Isfahan University of Thechnology Isfahan 84156 83111 Iran Degree M Sc Language Farsi Supervisor Dr Shadrokh Samavi samavi96@cc iut ac irAbstractPopulation of old generation is increasing in most countries Falling is one of the most dangerous events thatmay happen for these people and needs immediate medical care Automatic fall detection systems help themstay alone at home and reduce the burden on healthcare system Visual systems have advantage overwearable devices that do not disturb the normal life of the people They extract some features form videosequences and decide based on them Commonly used features have disadvantage of being view dependent Using several cameras to solve this problem increases the complexity of final system In this project weexploit a drawback of one simple background subtraction method and propose the variations in silhouettearea as a feature that is robust to view direction We use running average method for background subtractionand show experimentally and mathematically that variations in silhouette area can be a measure of rapidmotion during the fall impact to surrounding environment and inactivity after fall We use support vectormachines for classification and propose a scheme for combining several features The dataset under test is alarge one that has especially been created for fall detection algorithms and is publicly available Classification based on silhouette area produced error rate of 4 6 percent that is far better that commonly usedfeatures such as bounding box ratio and vertical velocity Most false classifications happen in sitting andlying that are wrongly classified as falls If we define the chair and sofa as inactivity zones the error rate ofthe system decreases to 2 1 percent These results are comparable to complex multi camera systems that usestate of the art features Generally in vision based systems the accuracy of the extracted features stronglydepends on the quality of silhouette that is never exact Our system does not need the exact silhouette but ituses the inaccuracy of silhouette to achieve its goal This makes the final system simple and a perfectcandidate for commercial implementation Our proposed method which has low computational burden hasthe potential of hardware realization inside camera The output of such a camera would be alarms orfeatures not videos that is very important to retain personal privacy Key wordsfall detection visual surveillance silhouette area view invariant support vector machines

شادرخ سماوي

بهزاد نظري، محمدرضا احمدزاده