method for recognizing the posture of a lactating sow through double-flow RGB-D Faster R-CNN

Abstract

The invention discloses a lactating sow posture recognition method based on double-flow RGB-D Fast R-CNN. The end-to-end double-flow RGB-D Faster R-CNN algorithm fusing RGB-D image features in the feature extraction stage is provided and used for recognizing five postures of standing, sitting, prone lying, abdominal lying and lateral lying of lactating sows in a free fence sow scene. Based on theFaster R-CNN, firstly, two CNN networks are used for extracting RGB image features and depth image features respectively; Generating an interested area of the RGB image feature map and the depth imagefeature map only by adopting one RPN network by utilizing a mapping relation of the RGB-D image; After pooling the features of the region of interest, realizing splicing fusion of RGB-D features by using an independent network layer; And finally, in the Fast R-CNN stage, introducing an NOC structure to continue to perform convolution extraction on the fused features, and then sending the featuresinto a classifier and a regression device. The invention provides a high-precision and small model fused with RGB-D data information end-to-end and a real-time sow posture recognition method, and a foundation is laid for further analysis of sow behaviors.

Description

technical field [0001] The present invention relates to the technical field of multi-modal target detection and recognition in computer vision, in particular to a Faster R-CNN based target detection algorithm, using RGB-D data, using dual-stream CNN to extract RGB-D features and then merging them in the feature extraction stage An end-to-end approach to lactating sow pose recognition. Background technique [0002] The behavior of pigs in pig farms is an important manifestation of their welfare and health status, which directly affects the economic benefits of pig farms. In the monitoring technology of animal behavior, compared with traditional manual monitoring and sensor technology, automatic recognition using computer vision is a low-cost, high-efficiency, non-contact way that can continuously provide valuable behavioral information. [0003] Behavior recognition in pigs using computer vision has been extensively studied in recent years. For example: In 2018, Xue Yueju o...

AI Technical Summary

Problems solved by technology

In addition, the depth image lacks the detailed information features of the target object due to the lack of information such as texture and color of the RGB image. It is difficult to accurately identify the target object when the shape is highly similar.

In particular, in the pose recognition task of sows taken from the top view, on the one hand, the height of the target is important information for judging different poses, which cannot be reflected in RGB images; on the other hand, the height and The shapes are similar (for example, lying prone and lying on the stomach), and it is difficult to accurately distinguish them using only depth information

Images