Research On Algorithms Of Video Analysis-based Fall Event Detection

Fall event detection is an important research topic in the field of computer vision.It has a very broad prospect in the field of smart home and public security.Whether at home or in public,a fall event detection method can promptly issue an alarm after a fall event is detected,which will earn valuable time for follow-up treatment.Therefore,it is of great significance to study the methods of fall events detection.This thesis focuses on the fall events detection based on video analysis under the framework of machine learning.First work is studying a fall events detection method in solitary scene.Due to the lack of fall detection dataset in crowd scene,the second work is constructing a fall events dataset in crowd scene.Finally,on the basis of the new data set,the research work on the detection method of fall events in crowd scene is carried out.The specific research work is as follows:After investigating a variety of falling detection methods,a method based on motion history image and histogram of oriented gradients is proposed.First,the faster R-CNN is used to detect the pedestrians on the input video images and map the bounding boxes to the corresponding motion history images,and then get the sub motion history images.And next,the HOG feature is extracted from each sub MHI.Finally,the support vector machine is used to classify the features,and the results of the detection are obtained.Compared with the present methods,this method effectively improves the recall,accuracy and the character of real-time of fall event detection.However,this method is suitable for the solitary scene with a fixed background,and it is difficult to detect the fall events in the crowded scene.For most crowded stampede accidents caused by fall events,and the lack of datasets in the study of fall event detection method in the crowed scene.A dataset of fall events in a crowed scene is constructed,and a fall detection method based on fully convolution network heatmap is proposed.The constructed dataset include six actions under three different scenes.After completing the constructing of the dataset,this thesis proposes a fall event detection method based on fully convolution network heatmap in crowd scenes.In this method,the bounding boxes are extracted with the heatmap generated by the full convolution network,and then the picture blocks in the bounding boxes are extracted.Then the picture blocks are put into the convolution network to extract the features and classify them,and the results of the detection are obtained.The experiment results show that the method based on the fully convolution network heat map outperform the compared method in this dataset.