The Research On Kinect-based Fall Detection

In recent years,with the improvement of living level,the development of computer and sensor techniques,and the increasing aging of the world population,the fall detection method has provided lots of convenience and cost savings for the elderly and patients under the medical supervision.According to the data acquisition,fall detection methods can be divided into two types: sensors-based and camera-based.The sensor-based fall detection method requires the object to wear some wearable sensor node,and let him/her move freely.While the camera-based method does not require the object to wear any equipment,which does not affect the movements of object.But the object should stay in the specific scenario.This thesis focuses on the camera-based fall detection,in which the 3D cameraKinect produced by Microsoft is used for fall detection.Different from the general 2D camera,it might improve the recognition accuracy and decrease the computation complexity with the 3D camera.The main works follows:To combat the problem of low recognition accuracy due to the feature extraction,in this thesis the eigenvectors are constructed during the feature selection phase by calculating the distances from the skeleton nodes to the ground,the acceleration of bone node and the distance between the points of K-means bone nodes.According to the variance of the skeleton nodes between the adjacent frames,the feature is captured to enrich the information.The feature vector is extracted from each sample,and it is trained and classified by HMM.The recognition accuracy and sensitivity are 95.5% and 100% respectively.Compared with the traditional fall detection methods,there are increased by 4.8% and 12.0% respectively.To combat the problem of low recognition accuracy due to the redundant information of bones,in this thesis,the block maps of variance are used to represent the discrete conditions of the 21 bone nodes for different actions.According to the variance transformed into the bone information,the means and standard deviations of the sliding window are transformed into the feature weights for the vectors.With the HMM training and classification,it further improves the recognition accuracy.The recognition accuracy reaches 96.6%.