A neural network-based fall detection method for multi-scale and multi-target

Abstract

The invention discloses a neural network-based fall detection method applicable to multiple scales and multiple targets, which mainly solves the problems that the existing fall detection algorithm has low detection accuracy for small targets and cannot realize multi-target detection. The plan is: (1) Supplement the public fall detection data set and create a fall detection fusion data set; (2) Transform the YOLOv4 network to build a YOLOv4 fall detection network for human characteristics; (3) use The K-means algorithm updates the anchors value for the fall detection fusion data set; (4) uses Label Smoothing to transform the network labels; (5) trains and tests the transformed YOLOv4 network. The invention improves the fall detection accuracy of small targets, realizes fall detection of multiple targets, can be applied to places where falls are prone to occur, and improves the rescue efficiency for fallen people.

Description

technical field [0001] The invention relates to the field of machine vision, in particular to a neural network-based fall detection method suitable for multi-scale and multi-target. Background technique [0002] In recent years, with the development of science and technology and the level of medical care, the aging of the population has continued to intensify. It often happens that the old man falls alone at home and no one finds out, and eventually dies. Therefore, the application of machine vision in fall detection has very important practical significance. [0003] The fall detection methods based on machine vision mainly include: the fall detection method based on traditional image processing and the fall detection method based on deep learning. The fall detection method based on traditional image processing is greatly affected by the detection environment and cannot be realized. Multi-target detection; the fall detection method based on deep learning is less affected b...

AI Technical Summary

Problems solved by technology

[0003] The fall detection methods based on machine vision mainly include: the fall detection method based on traditional image processing and the fall detection method based on deep learning. The fall detection method based on traditional image processing is greatly affected by the detection environment and cannot be realized. Multi-target detection; the fall detection method based on deep learning is less affected by the detection environment, and can accurately detect multiple targets

[0004] The published fall detection methods based on deep learning generally divide the detection process into two steps: moving object detection and fall judgment, using two networks for detection, a feature extraction in the moving object detection stage, and a fall judgment stage Perform feature extraction again, repeating feature extraction twice will greatly affect the detection speed, making it difficult to train the network model

In addition, the published target detection network model has low detection accuracy in complex backgrounds, does not perform well in extracting human signs, and lacks semantic information

The number of samples containing small targets in the public human behavior data set is very small, potentially making the target detection model pay more attention to large target detection, which will result in lower accuracy of small target detection

Images