Système de détection des chutes en ligne

Abstract

Self-injury and suicide in Canadian correctional facilities is a serious problem. This type of incident has increased in recent years and requires prevention and detection techniques to provide immediate intervention and reduce its impact. The most well known methods of suicide are overdose, hanging and cutting the wrist.This work focuses on the detection of the falls following an overdose suicide attempt. Our goal is to develop a method that detects the fall in real time in an efficient and reliable manner with a minimal false alarm rate. Most of the detection methods proposed in the literature use traditional techniques that limit the reliability and generality of the system. Additionally, these methods don’t work in real time, they only accept well segmented videos that contain only one action. To achieve our goal, we used the Kinect v2 from Microsoft which is an RGB-D camera. This sensor offers different data streams : RGB image, infrared, point cloud and human skeletal data. We chose to use the human skeleton to detect the fall since it is independent of the environment and lighting conditions. Four public databases were used to train our system : the TSTv2 and FallFree which are fall detection oriented databases ; plus the NTU-RGB+D and PKU-MMD which are action detection and recognition oriented databases. Our method is composed of two modules : an action detector and a fall detector. The first module is responsible for detecting the beginning and the end of each action. This module is based on the deep learning algorithm 1D-CNN. It cuts the continuous flow of skeletal data into subsequences of 15 images and classifies each subsequence independently. If the algorithm detects motion in the subsequence, the subsequence is classified as positive. The succession of positively classified subsequences forms an action. This module was trained with the PKU-MMD database. After extraction of the action, the role of the second module is to classify it as ’falling’ or ’not falling’. This module is based on a ST-GCN. It has the advantage of simultaneously capturing spatial and temporal information. It exploits the fact that the skeletons are in the form of graphs instead of 2D or 3D grids. First it was trained with the NTU-RGB+D base. Then, the transfer learning technique was applied using both TSTv2 and FallFree databases. The performance obtained exceeds that of the methods proposed in the literature on both TSTv2 and FallFree datasets.

Date	14 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Rita Noumeir (Supervisor)