In the naked eye 3D video viewing, due to the contradiction of convergence adjustment, viewers will have visual fatigue symptoms such as dizziness and vomiting. Especially in medical, A/VR and other scenes, 3D video viewing is usually driven by special task. As eye movement behavior is subjectively controlled, it is difficult to accurately assess visual fatigue with a single eye movement feature, let alone obtain the visual fatigue level, resulting in visual fatigue not being detected in time and causing unpredictable losses. To solve the above problems, this paper proposes a 3D visual fatigue rating model based on a variety of eye movements, aiming to obtain the real-time visual fatigue rating by modeling and analyzing the eye movements such as fixation, scanning and blinking when viewing 3D content. The method combining subjective and objective is used to carry out the experiment: the subjects watch 3D content, the experimenters record subjective scores of the subjects, and extract the objective eye movement behavior of the subjects with an eye tracker. Through correlation analysis, the various objective eye movement behaviors that can represent 3D visual fatigue under task driving are explored. A four grade evaluation model of visual fatigue based on 16 eye movements is established by using neural network. The prediction accuracy of the model for 3D visual fatigue level reaches 82%, proving the effectiveness of the model.