The work aims to propose a collision detection method combining AABB enclosing box coarse detection and depth map fine detection, in order to improve the efficiency of collision detection in the virtual assembly system. First of all, SolidWorks software was used to build a model of the BZ350 candy packaging machine and import it into the engine in STL format. An AABB enclosing box was built for the part to carry out coarse detection, and then the corresponding depth maps for the two parts were obtained respectively, and the two depth maps were superposed with OpenCV. Next, according to the superimposed pixel values, whether the parts collided or not was judged. The experiment showed that when the number of superimposed depth values less than 100 accounted for more than 86.2736% of all depth values other than 0, it could be recognized that the two parts collided. Compared with the traditional detection algorithm, this algorithm has advantages for the detection of complex models under the premise of guaranteeing the detection speed and detection accuracy and ensuring that the running speed is above 40 frames/s to run smoothly.