The traditional algorithms of collision detection often have poor execution efficiency in dealing with the aggregation following collision of a large number of particles. Therefore, a two-step, fast and accurate algorithm of collision detection was proposed based on the pretreatment of bounding sphere and maximum detection region. In the broad phase, all aggregates were represented by bounding spheres with low update cost. The detection of collisions between bounding spheres was converted into solving the problem of quadratic equations regarding time, and all possible collisions were detected fast by solving these equations parallelly. The detection of collisions outside the maximum detection region was ignored to further reduce execution time. In the narrow phase, the specific time and position of collisions were rapidly determined by discrete collision detection, where sampling time intervals were self-adaptive and decreasing. Simulation results were compared with those by non-optimized traditional algorithms, and it was found that, on the premise of meeting the same accuracy of collision detection the algorithm proposed here could increase the execution efficiency 10 to 30 times, indicating that this algorithm is more applicable to collision detection during the aggregation process of a large number of particles.