To improve the efficiency of collision detection in complex scenes, a collision detection algorithm based on topological spatial grid is proposed in this paper. Because there are numerous objects with complex shapes, different scales and motion states in the scene, the algorithm firstly partitions the scene with uniform octree, builds oriented bounding box hierarchal tree, sets up the topology of spatial grids and conducts static large-scale object decomposition for accurate positioning in preprocess stage. Then, the algorithm excludes plenty of non-intersect objects with topological spatial grid and projection intersection test, by using bounding volume hierarchy to conduct accurate detection of potential collision pairs and calculates collision points in real-time detection process. Experimental results show that this algorithm improves the real-time detection efficiency, which is suitable for collision detection in complex virtual scenes.