With the construction of modern sheep farms, it is of great significance to achieve accurate and real-time detection as well as counting on sheep. In order to solve the problem of overlapping occlusion among sheep group and chase the need for real-time performance, this paper proposes a Light Attention YOLO model, which fuses the attention depth separable module with CSPDarknet and PAFPN respectively, adds CBAM attention module to the 4th and 5th layers of the backbone. The model is trained and tested using 1225 monitor frame images. Through completely testing in the hardware environment of a single GTX1080T¡ graphics card, the mean accuracy precision of this model in different environments reaches 92%, and the detection speed is 28.5 frames per second, which perfectly meets the accuracy and real-time requirements. In different complex environments, the comprehensive performance on sheep detection is better than the current mainstream detection model. The model can be deployed on the web system to achieve real-time sheep detection and counting.