Research On Human Fall Detection With Complex Backgrounds

With the acceleration of the aging population, the phenomenon of the aged people living alone is increasing popular. If the elderly occur accidental falls at home, it is likely to cause serious injury without timely treatment, even death, so the research on fall detection is of great social significance and has become the current research focus at home and abroad. Taking an overview of the study about fall detection at home and abroad, most of them are based on video and image sequence analysis, while the process needs a large amount of image and data processing, resulting in a large amount of information redundancy, thus those methods can hardly meet the real-time requirements. Also if the background is too complex, the accuracy rate will be severely reduced.In order to solve the above problem, this paper proposes a fall detection method based on multi-feature fusion under complex background, which is designed to detect the fall action more quickly and accurately, so it is more likely to gain much time for the subsequent treatment and reduce the possibility of disability. The main contents of this paper include:1. The establishment of human probability model. This model has laid an important foundation for the follow-up process to segment the human body from the static image. Because of uncertainty and complexity of human body posture and the image background, it is unrealistic to segment the whole body out at one time. The idea of this method is firstly to segment the image into multiple regions; then calculate the probability of each region contains parts of the human body, or the probability of the region is part of the human body; finally calculated again that the possibility for all the of each region contains parts of the human body. If greater than the set threshold, the human body part can be ensured and then segment the human body combined with the segmentation method.2. Research on the method of human body segmentation in a static image. Because of the complex background, the general methods cannot segment human body from the complex background very well. A fusion depth human body segmentation algorithm is proposed in this paper, which is the fusion of watershed segmentation algorithm and automatic region growing algorithm. Watershed segmentation algorithm can segment the regions with similar characteristics perfectly, but it is easy to produce the over segmentation problem, while automatic region growing algorithm can solve the problem by autonomic selection of seed region, some similar areas but belong to different ones can be combined into one area. The experimental results show that the algorithm works well.3. According to the characteristics of the fall posture selection problem, this paper proposed a fall feature vector model. This paper redefines the feature vector model is used to detect fall action, contains a total of5categories of11feature vectors, the five categories are the human aspect ratio, the effective area ratio of human body, the point-edge distance of human body, the axis angle of human body and the centrifugal rate of human body. The purpose of using this feature vector model is training support vector machine which is based on Gauss Radial Basis Function, which the feature space is infinite dimensional Hilbert space, the maximum interval optimization problem even can be solved in infinite dimensions, which can meet the demands of this paper.The simulation results show that, the human body segmentation method in this paper has good applicability to complex background, the accuracy of the segmentation reached96.15%and the time is72.2ms. Compared to other image segmentation algorithms, this method has better segmentation effect and less processing time; The support vector machine classification model based on Gauss Radial Basis Function has a good static image accuracy rate reaching94.5%, and the recognition accuracy of the video image sequence is even higher, reaching94.9%. This study proves that the model can satisfy the demand of fall detection in complex background with rapid speed and high accuracy, which is of great practical significance to improve the life quality of the elderly.