Behavior recognition method based on single stage

Abstract

The invention discloses a behavior recognition method based on a single stage. A feature pyramid network structure comprises a bottom-up structure on the left side, a top-down structure on the right side, a transverse connection layer located in the middle and a prediction network located at the rear end; the bottom-up structure on the left side comprises a first convolution layer, a second convolution layer, a third convolution layer, an RFB module and an RFB-s module; the top-down structure on the right side comprises a fourth convolution layer, a fifth convolution layer and a sixth convolution layer; the bottom-up structure is used for continuously extracting feature information; and the top-down structure on the right side continuously amplifies the feature map of the top layer throughdown-sampling, namely deconvolution, and the transverse connection layer combines the features of the top layer on the right side with the features of the bottom layer on the left side, so that the features have better expression ability, and the prediction network is used for realizing classification and prediction of a bounding box.

Description

technical field [0001] The invention relates to the fields of computer vision, target detection and image processing, in particular to a single-stage behavior recognition method. Background technique [0002] With the continuous development of deep learning, good progress has also been made in natural language processing and object detection and classification. Among them, the target detection is mainly divided into two branches: one branch is the target detection based on the region proposal, such as: RCNN series (RCNN, Fast RCNN and RFCN, etc.), these target detection methods are mainly divided into two stages. In the first stage, high-quality candidate boxes are generated through the algorithm or the region proposal network, and then in the second stage, these candidate boxes are classified and bounded by the sub-network. Therefore, since this type of detection method is carried out in two stages There are defects in speed, and it cannot achieve real-time effects; the ot...

AI Technical Summary

Problems solved by technology

Finally, Focal loss is introduced in the prediction stage to solve the imbalance between positive and negative samples. By combining the three excellent methods, the detection accuracy of the method can be improved on the basis of ensuring real-time detection.

Images