In view of harvesting damages and failures caused by the absence of attitude information in the current apple harvesting robot, the method of measuring apple attitude was put forward with the machine vision, and the optimal estimation of apple attitude was carried out based on the particle filter in real time. A camera was fixed on the robotic gripper, in the process of picking the camera vision field was controlled to cover the expected fruit target, and a series of apple images were sampled randomly. The algorithms of the inertial axis and the center offset were applied to calculate the apple attitude at the robotic coordinate according to the each apple image. Finally, the attitude data extracted from many apple images was fused with the particle filter, and the optimal estimation of the apple attitude was obtained. The experimental results showed that the multiple information of apple attitude was fused successfully, and the measurement error of apple attitude was decreased effectively.