E3D: An efficient 3D CNN for the recognition of dairy cow's basic motion behavior

Accurately and rapidly recognizing the basic motion behaviors (lying, standing, walking, drinking, and feeding) is helpful in better understanding the health status of dairy cows. Existing algorithms cannot effectively deal with the problem of large parameters, thus difficult to load and use on portable edge devices. In this paper, an E3D (Efficient 3D CNN) algorithm was proposed to solve the problems of existing algorithms. Based on the 3D convolution combined with Dwise (Depthwise Separable Convolution) in the SandGlass-3D module, E3D could directly and efficiently process the Spatial-Temporal information of the video. The ECA (Efficient Channel Attention) was introduced to filter channel information for accuracy improvement. Experimental results showed that the precision, recall, parameters, and FLOPs of the E3D were 98.17 %, 97.08 %, 2.35 M, and 0.98 G, respectively. The accuracy of E3D was 7.29 %, 4.06 %, 5.31 %, and 12.46 % higher than C3D, I3D, P3D, and S3D, respectively. The parameters were reduced by 11.95 M, 25.73 M, and 280.65 M compared with the Improved Renext network, ACTION-Net, and C3D-ConvLSTM. It indicated that the proposed network was suitable for accurately and rapidly recognizing the basic motion behaviors of dairy cows in natural environments.

Automated summary

In this paper, the E3D algorithm was proposed to accurately and rapidly recognize the basic motion behaviors of dairy cows. The algorithm combines 3D convolution and Dwise to efficiently process the spatial-temporal information of the video. Experimental results showed that E3D outperformed other networks in terms of accuracy and parameter reduction, making it suitable for recognizing cow behaviors in natural environments.