Action Classification Method Based on Joint Spatiotemporal Simple Recurrent Network and Attention Mechanism

Abstract

The invention relates to an action classification method based on joint point space-time simple loop network and attention mechanism. First, the ordinary SRU model is extended to the space dimension, and a ST-SRU model that performs iterative calculations in both time and space dimensions is designed. Then, based on the proposed ST‑SRU model, the global context attention mechanism is introduced, and the GCA‑ST‑SRU method is proposed. Finally, the proposed method is used for human behavior recognition. First, the features of human joints are extracted by deep network, and then the extracted features are recognized by GCA‑ST‑SRU method. The method of the invention can reduce training time consumption and improve classification accuracy, and has obvious efficiency advantages. The rapid inference speed of the method of the invention is beneficial to the design of a real-time action recognition system, is suitable for running on a platform with limited computing power, and has broad application prospects in the fields of computer vision, intelligent monitoring, human-computer interaction and the like.

Description

technical field [0001] The invention belongs to the field of pattern recognition, and is a method for modeling actions represented by joint points by using a space-time simple loop network, and combining the advantages of an attention mechanism to classify actions. Background technique [0002] Action recognition is widely used in intelligent video surveillance, human-computer interaction, medical assistance, abnormal behavior detection and other fields. Action recognition refers to the classification of the behavior of characters in a video. A video can be decomposed into multiple consecutive pictures, so action recognition can also be regarded as a problem of classifying a sequence of pictures. A popular research direction of action recognition in recent years is to capture the three-dimensional space coordinates of body joint points when people are doing actions through sensors such as depth cameras, and then classify the sequence of human joint point coordinates. The coo...

AI Technical Summary

Problems solved by technology

[0006] The above methods have made good use of various recurrent network models to conduct research on behavior recognition based on human joints, but methods based on LSTM and GRU have a large number of calculations that depend on time series when processing sequences of human joints, and these calculations cannot finish independently

For example, when LSTM and GRU calculate the hidden state of the current time step, due to the dependence, they must first calculate the hidden state of the previous time step, which limits the speed of sequence processing. With the scale of the LSTM model and the number of hyperparameters The increase of the training time is getting longer and longer, and the adjustment of the parameters will also take more time

Images