In view of the problem that the extraction accuracy of beef cattle skeleton was decreased greatly with the increase of targets in the process of beef cattle behavior recognition, an improved YOLO v3 algorithm (Not classify RFB-YOLO v3, NC-YOLO v3) was proposed. After the backbone network, receptive field block (RFB) was introduced to expand the receptive field of the model, and the classification module was eliminated to improve the detection efficiency. Combining 8SH (8-Stacked Hourglass) algorithm to realize multi-target detection and skeleton extraction of beef cattle in actual breeding environment. In the experiment, totally 16 key nodes were set for the beef cattle skeleton to form the beef cattle pose point information, and the detection accuracy was improved through multi-scale and multi-direction training of the image. Based on the statistical analysis of key points of multi-target skeleton extraction model, a method for beef cattle standing and lying down behavior recognition was proposed. Experimental results showed that in terms of target detection, the recall of the NC-YOLO v3 model can reach 99.00%, the precision can reach 97.80%, and the average precision can reach 97.18%. Compared with the original model, average precision was increased by 4.13 percentage points, and the amount of network parameters removed was 13.81MB;in terms of single-ox skeleton extraction, the 8-Stacked Hourglass network was used to detect key point positions, and the mean average precision can reach 90.75%. In terms of multi cattle skeleton extraction, compared with the model constructed by YOLO v3, the mean average precision of the model constructed by NC-YOLO v3 was increased by 4.11 percentage points to 66.05%.