Behavior identification method based on space-time context association

Abstract

The invention belongs to the technical field of identification, provides a behavior identification method of a deep network model based on space-time context association, and aims to solve the problemthat the spatial feature learning range in a CNN model is limited by the size of a sensing domain, reduce the loss of behavior feature representation by the model, and improve the accuracy of behavior recognition. According to the main scheme, user behavior data is imported for convolution mapping operation; then a behavior space-time characteristic graph TSF is obtained by utilizing a grid LSTMneural network, the behavior space-time characteristic graph TSF is imported into an attention gate module to carry out different time characteristic weight learning, a behavior characteristic graph is obtained, and the behavior characteristic graph is input into a softmax classifier to calculate probability distribution D of a behavior category; cross entropy loss function operation is carried out on the probability distribution D of the behavior category and the training set behavior label Y to obtain a loss Loss0, and an l2 loss function is introduced as a final total loss function L; and according to the total loss function L, the numerical value of the model imaginary-seating parameter is modified by using a back propagation operation to obtain the deep network model M.

Description

technical field [0001] The invention belongs to the technical field of deep learning, pattern classification and recognition, and in particular relates to a behavior recognition method based on a deep network model associated with spatio-temporal context. Background technique [0002] Deep learning methods bring strong feature extraction capabilities to behavior (mainly divided into periodic behaviors (such as walking, running, cycling, etc.) and some scattered behaviors (such as drinking tea, taking cups, cooking, etc.)) , which reduces the difficulty of artificial feature modeling. Behavioral data based on multi-sensors is not only time-sequential, but also has certain local spatial characteristics between sensors. For the spatio-temporal characteristics of behavior, there are certain defects in the current deep learning-based methods. For example, although the RNN-based method can learn long-term feature dependencies, it cannot describe local features well. However, CNN...

AI Technical Summary

Problems solved by technology

For the spatiotemporal characteristics of behavior, there are certain defects in the current deep learning-based methods. For example, although the RNN-based method can learn long-term feature dependencies, it cannot describe local features well.

However, CNN-based methods are not ideal for long-term correlation feature extraction.

Although the CNN+RNN method can further improve the learning ability of spatio-temporal features, the method of first using CNN to extract spatial features from data and then performing time-related modeling damages the potential context-dependent characteristics of time-series data to a certain extent.

In addition, the model is also very bloated, which is difficult to apply to wearable devices with limited computing power and battery capacity

Images