Behavior identification method combined with space-time discrimination filter bank

Abstract

The invention discloses a behavior identification method combined with a space-time discrimination filter bank. The method comprises the following steps: constructing a global branch which comprises aconvolution layer, a global average pool layer and a linear classifier, expressing the classification output as Zavg, aggregating the whole spatio-temporal information from a video by the branch through average pooling operation, and capturing some significant feature differences in the video; constructing a local detail branch, introducing a group of filters as a classifier Zmax of a local area,introducing a cross-channel pooling layer and a softmax loss layer, and directly performing supervising on the filters to construct a classifier Zchannel; and combining the output Zavg, the output Zchannel and the output Zmax of the three classifiers into a single prediction Z. According to the method, the network learning capability can be effectively improved. According to the method, a large amount of calculation amount cannot be increased, and the recognition capability of the network for complex behaviors can be effectively improved.

Description

Technical field [0001] The invention relates to a behavior recognition method, in particular to a behavior recognition method combined with a spatio-temporal discrimination filter bank. Background technique [0002] In recent years, with the development of society, various electronic products have emerged one after another, and a large number of videos are produced every day. Video data plays an important role in information transmission in daily life. It is very important to analyze and understand the massive video content. significance. The video data contains a large number of video clips based on human related behaviors. Video-based human behavior recognition can automatically identify the human behavior in the video. In video surveillance, human-computer interaction, sports video analysis and video retrieval, etc. The field has a wide range of applications. Therefore, video-based human behavior recognition has always been an active topic in the field of computer vision, wit...

AI Technical Summary

Problems solved by technology

[0005] Similar to image classification, the existing neural networks applied to video behavior recognition have similar last few layers, a linear classifier composed of a global average pooling layer followed by a fully connected layer, but such a network architecture has a significant impact on Different actions (for example, different backgrounds, actions with large differences, etc.) have better results, but it cannot capture the fine discriminative information in the video, so for some complex actions with similar actions but different detailed information cannot be obtained very good classification effect

Images