Basic motion behavior recognition of single dairy cow based on improved Rexnet 3D network

Abstract

• Rexnet was extended to Rexnet 3D for basic motion behavior recognition of dairy cows. • Sliding window mechanism were used to verify the effectiveness of untrimmed videos. • Rexnet 3D could be used for recognizing the basic motion behavior of dairy cows. Basic motion behaviors of dairy cows (lying, standing and walking) contain abundant health information. However, to monitor these behaviors accurately and efficiently by using camera-based methods remains challenging. In this study, time dimension was added to extend the Rexnet to Rank eXpansion Network 3D algorithm (Rexnet 3D) network to realize non-contact and automatic recognition of dairy cows’ basic motion behaviors. A total of 406 videos containing 256,500 frames of dairy cows in different scenes and postures were collected and tested. Six different networks including Resnet101, Mobilev2, Mobilev3, Shufflev2, C3D and S3D were selected to validate the performance of the proposed network. The average number of frames recognized per second (ANFR), FLOPs and Networks Weight were used to evaluate the performance of the method. Experimental results showed that behavior recognition accuracy, ANFR, FLOPs and Networks Weight of the proposed method were 95.00%, 76.52 fps, 15.80 G and 14.30 M, respectively. Meanwhile, a total of 10 unedited videos with 1,722 s in total (43,065 frames) were collected to verify the effectiveness in real environment. The results showed that the accuracy of the proposed method was higher when the sampling interval was five frames and the accuracy of the specific three basic motion behaviors was 91.02%, and the ANFR was 101.02 fps. In general, these findings indicated that it was feasible to use a designed network for accurate recognition of cows’ basic motion behaviors. The network could provide necessary reference for improving the precision breeding and health welfare of dairy cows.