An Adaptive Temporal Shift Neural Network Timing Behavior Recognition Method

Abstract

The present invention proposes a sequential behavior recognition method of an adaptive sequential shift neural network, which firstly collects and models features of multiple time points; then introduces an adaptive sequential shift neural network to learn the receptive field required by each layer of the network; Finally, the learnable displacement variable is trained to correct the bone point data. The present invention can adaptively learn the receptive field required by each layer of network, and can adaptively learn the receptive field required by each data set. The adaptive temporal shift neural network can learn different time shift vectors for different data, so as to adaptively adapt to different data distributions. Through the timing behavior recognition method proposed by the present invention, it is possible to save computing resources while improving behavior detection accuracy. This kind of adaptive learning is more superior than manual parameter adjustment of ordinary time convolution.

Description

technical field [0001] The invention relates to an adaptive time sequence shift neural network time sequence behavior recognition method, relates to the general image data processing or G06T generation field, and particularly relates to the G06T 7 / 20 motion analysis field. Background technique [0002] With the development of artificial intelligence, the recognition of human behavior has received more and more attention. Recognition of human behavior can be used in security, human-computer interaction and other fields. [0003] In the research of behavior recognition, a hot issue is how to perform time-series behavior recognition. The so-called time series behavior refers to the behavior that cannot be judged by a single frame of image, but needs to be judged by observing a time sequence action. For example, standing up and sitting down, these two behaviors are difficult to distinguish through a single frame of image, and can only be distinguished by observing a time serie...

AI Technical Summary

Problems solved by technology

[0004] The existing time series shifting method uses a time series shift module to model the features at different time points. For the time series convolution and time series shift methods, the size of the receptive field is artificially specified, and this The artificially specified receptive field is not suitable for temporal behavior recognition tasks; in temporal behavior recognition, different databases require different receptive fields, which leads to the need for a large number of parameter tuning experiments on different databases

Images