Ostrich behavior is a key sign of their development and health. Quickly and accurately identifying it is crucial for growth monitoring and disease prevention. Computer vision technology, being real-time and non-contact, is widely used in livestock behavior recognition. But current methods, mainly for common livestock in simple settings, have drawbacks.This paper presents a method for ostrich behavior recognition using YOLOv7-MG. It aims to boost recognition efficiency and precision. Images of ostrich behavior are gathered from actual farms to create a dataset. The MobileOne network replaces the backbone of YOLOv7 to cut down computation and model size. Also, a GAM module is added to improve feature extraction in complex situations.The proposed method does better than YOLOv7 and other cattle behavior recognition systems. It has a relatively small model memory footprint and can precisely identify ostrich behavior. This lays the groundwork for ostrich disease prevention and management.