Fall behavior detection method and device based on depth space-time convolution auto-encoder

Abstract

The invention discloses a fall behavior detection method and device based on a depth space-time convolution auto-encoder, and belongs to the technical field of digital image recognition, and the method comprises the steps: preprocessing video frame data of normal activity; carrying out noise reduction processing on the data after preprocessing the normally active video frame data; adopting 3D convolution to construct a deep space-time convolution auto-encoder fall anomaly detection model; training a deep space-time convolution auto-encoder tumble anomaly detection model; and identifying that the video frame has a falling abnormity. The device comprises a video frame acquisition module, a preprocessing module, a noise reduction module, a construction module, a training module and an identification module. fall detection is regarded as an abnormal detection problem, and due to the fact that fall is rare in daily life, sample categories are imbalanced, Therefore, by applying anomaly detection to the falling problem, the category of the sample can be balanced, and a better detection effect can be displayed.

Description

technical field [0001] The invention belongs to the technical field of digital image recognition, and in particular relates to a fall behavior detection method and device based on a deep spatio-temporal convolutional self-encoder. Background technique [0002] With the continuous development of society, population aging has become an inevitable problem in the development of our country and the world. Statistics show that among the factors that threaten the health of the elderly, falls cause the largest proportion of injuries, and the resulting medical costs are quite huge. Detecting the occurrence of falls in time and enabling the elderly to receive timely medical assistance can directly and effectively reduce the risk of death caused by falls. [0003] At this stage, human fall detection is mainly divided into two major research methods: sensor-based methods and computer vision-based methods. Sensor-based methods generally collect data through wearable devices and provide...

AI Technical Summary

Problems solved by technology

Although the above method can detect whether a fall has occurred based on certain characteristics of the fall, since the fall detection method based on RGB video images needs to install a camera in the monitoring area, and detect whether a fall occurs through image and video information, this method will have a negative impact on The privacy of the human body poses certain threats and violations, so it generally cannot be applied to privacy-sensitive spatial areas

Images