Abstract: To achieve fast and accurate detection of livestock targets in grazing areas, we proposed a lightweight livestock target detection algorithm with improved YOLOV3-tiny, which is a lightweight object detection algorithm for real-time detection on Jetson Nano embedded motherboard. In terms of network structure, the anchor frame clustering algorithm is optimized according to the characteristics of livestock targets in grazing areas, and the prediction output scale is increased to enhance the use of shallow information. The pyramid network is used for multi-scale feature fusion to improve the detection rate of small targets while ensuring the detection rate of large targets. The improved target detection mechanism can effectively improve the accuracy of target detection under complex light conditions (e.g., direct sunlight). The experimental results showed that the detection accuracy of the improved YOLOV3-tiny algorithm reached 83.2%, and the detection speed on the embedded platform Jetson Nano was 12 frames·s^-1. The algorithm improved the detection accuracy by 8.7% on average compared with the YOLOV3-tiny algorithm while satisfying the portability.

Key words: convolutional neural network, YOLO network, deep learning, smart pastoral.