To improve the accuracy and efficiency of classic AABB collide detection, An algorithm of the virtual sensors which is applied to the domain of Rover manipulator collide detection is proposes. At first, the inverse kinematics model of rover manipulator is established. Secondly, lunar terrain point cloud data is triangulated by Delaunay triangulation method, the triangulation face set is then stored in the leaf nodes of a multiple tree (AABB tree). In the end, the rover manipulator model is simplified by virtual sensors. By utilizing the virtual sensors to traverse the face set stored in the multiple tree, manipulator/lunar terrain collision are detected and avoided. The inner-yard experiment of lunar rover in-situ exploration task show that the accuracy of the collide detection increased to within 1mm and the time of collide detection decreased to 10 s by the algorithm proposed. The feasibility and efficiency of the collide detection algorithm is justified.