Multi-level deep recursion network group behavior identification method based on context

Abstract

The invention discloses a multi-level deep recursion network group behavior identification method based on context. The method comprises the following steps: step S1, encoding sub-action information through a binaryzation encoder based on the context, thereby forming single dynamic information; step S2, serving the movement locus of the human produced by using a human body detection and tracking method as the single individual locus, dividing all single individual locus into the human body group with time-space consistency, and establishing a single interaction model by using the single dynamic information so as to perform modelling on the human body group internal interaction and the interaction between the human body groups; and step S3, training the proposed multi-level recursion context encoding network so as to learn the features of the single dynamic information, the human body group internal interaction and the interaction between the human body groups. The method proposed by the invention is effective for identifying the group behavior, has robustness for the human body detection under the complex environment, and has enough high flexibility to simulate the high-order interaction context situation.

Description

Technical field [0001] The invention relates to a method in the technical field of group behavior recognition, in particular to a context-based multi-level deep recurrent neural network group behavior recognition method. Background technique [0002] High-level interactive context modeling, such as group interaction, is the core of group behavior recognition. However, most of the previous behavior recognition methods do not provide a flexible and extensible method to solve high-level context modeling problems. [0003] Group behavior recognition provides useful information for many practical applications including role understanding in the scene and event prediction. The main challenge of group behavior recognition is to model the interactive context information in the crowd. This is because the number of characters included in a scene often varies. In addition, in most cases, group behavior is related to several sub-interaction scenarios, so how to model the interaction between...

AI Technical Summary

Problems solved by technology

This is very helpful to solve the situation where the characteristics of the input instance are variable (for example, the number of characters in the group is different or the number of groups is different), and the linear expansion of the high-order context modeling problem is very flexible, which fundamentally solves the existing behavior recognition. approach does not provide a flexible, scalable approach to address the bottleneck of high-order context modeling

Images