The efficient visibility culling method is one of the important research aspects of real time rendering of complexle dynamic scene. This paper researches the visibility culling problem and improves the coherent hierarchical culling (CHC) algorithm. To handle the redundancy and unnecessary visibility culling of CHC, a probability computing model is presented. The model first calculates the expectation of occlusion query time and rendering time, then compares them to improve the query strategy of CHC algorithm. Experimental results show that with the model high culling efficiency can be achieved in the dynamic scene with highly complex depth and large amount of objects, and also achieve real-time rendering results.