Høgskolen i Gjøvik Saleh Alaliyat Video - based Fall Detection in Elderly's Houses.
2 Outline Introduction Background Proposed System Implementations Conclusion
Video - based Fall Detection in Elderly's Houses. 3 Introduction
Video - based Fall Detection in Elderly's Houses. 4 Objective and benefits: The main goal of this project is to detect person falling event in elderly’s houses and give an alarm in real-time. Ensure the safety of elderly people: Fast growing population of seniors. Shortage of employees taking care of seniors. The majority of injury-related hospitalizations for seniors result from falls.
Video - based Fall Detection in Elderly's Houses. 5 Background
Video - based Fall Detection in Elderly's Houses. 6 Background 1 Fall Detection techniques: Sensors wearable sensors. Infrared sensors (vertical velocity). Drawbacks: forget to wear them and not sufficient to discriminate a fall from sitting. Video – based mehtods
Video - based Fall Detection in Elderly's Houses. 7 Background 2 Indoor Surveillance Segmentation and Tracking Features extraction Events Classification
Video - based Fall Detection in Elderly's Houses. 8 Proposed System
Video - based Fall Detection in Elderly's Houses. 9
10 Implementations
Video - based Fall Detection in Elderly's Houses. 11 Segmentation The aim is to have a foreground image that has only the moving objects.
Video - based Fall Detection in Elderly's Houses. 12 a: Background Reference b: Current Frame c: Absolute difference d: Binary Image e: shadow mask f: Binary Improved
Video - based Fall Detection in Elderly's Houses. 13 Features extraction
Video - based Fall Detection in Elderly's Houses. 14 Applying median filter for all the extracted features for smothing. Motion before using median Filter. Motion after using median (window = 13) for smothing.
Video - based Fall Detection in Elderly's Houses. 15 Aspect Ratio: using X-Y Projections method (projecting the foreground pixels onto x and y axises). Aspect Ratio = Height / Width.
Video - based Fall Detection in Elderly's Houses. 16 Orientation: The angle between the x-axis and the major axis of the ellipse that represent the blob
Video - based Fall Detection in Elderly's Houses. 17 Motion Quantity: Sum of the pixels that belong to the blob and moving. Speed: the distance between the CoMs of the blob in a sequence of frames and divide it by the time. Height of the CoM: the distance between the CoM of the person and the floor.
Video - based Fall Detection in Elderly's Houses. 18 Vertical direction of the center of mass. MHI: Sum of the pixels values in the Motion History Image divided by the number of blob pixels.
Video - based Fall Detection in Elderly's Houses. 19 Audio: Audia signal Wavelet coefficients
Video - based Fall Detection in Elderly's Houses. 20 Sample window = 500; SNR: an indication of the difference in signal intesity. Test1: TV + talk + fall Test2: Music (song) + fall Test3: silence + Fall Test1 Test2 Test3
Video - based Fall Detection in Elderly's Houses. 21 Events Classification
Video - based Fall Detection in Elderly's Houses. 22 K-NN: the activities are classified in groups, walking and standing, sitting, and lying down. 24 short training movies (corridor in A-building and room A128. the movies have walking, standing, sitting, kneeing and falling (lying down). Make from them a trainng set for K-NN classifier (672). Test the K-NN by applying two test movies. Test1: 207 frames Start falling at frame #62 Full falling (lying down at frame #77 Stay lying down for 21 frames. K-NN output is lying down for these 21 frames.
Video - based Fall Detection in Elderly's Houses. 23 MHI : frame # 82; (after 3 frames from lying). Direction of motion: frame #68 (after 6 frames from fall starting). K-NN output is sitting (start giving Lying down at fame #72 to frame #117). Check the speed and motion quantity for next 45 frames if the object still in the lying position. Speed ( frame #82 to #105 = 23 frames).
Video - based Fall Detection in Elderly's Houses. 24 Conclusion
Video - based Fall Detection in Elderly's Houses. 25 Conclusion: K-NN gives confident results. including the audio. Future works: Define normal inactivity zones. Personal Information. 3D information.
Video - based Fall Detection in Elderly's Houses. 26 The end Thank you