Abstract: Behavior recognition is of great significance in cattle production and health management．Currently, there are some problems in cattle behavior recognition, such as high complexity of model and low accuracy in multi-objective recognition．To solve this problem, a lightweight multi-objective cattle behavior recognition model EVH-YOLO11 was proposed．In order to solve complex interference problems such as occlusion, overlap and small targets in actual cattle farm environment, a sandwich layout design was introduced for EfficientViT feature extraction module to reduce calculation redundancy．At same time, combined with Dynamic Head detection head, objective detection performance was adaptively enhanced．Experiments showed that EVH-YOLO11 outperformed mainstream models, which could provide technical support for smart cattle farms．