Accurately detecting the presence and evolving boundaries of cracks on rock surfaces is critical for understanding the behavior of crack evolutions and facture mechanism of rock and rock-like material, which could cause engineering disasters if proper operation were not taken to deal with the evolving cracks. In this paper, we investigate the problem of vision-based automatic detection of cracks on rock surface at pixel-level, which is a preliminary step of crack evolution analysis. We build a Split Hopkinson Pressure Bar (SHPB) system to simulate the crack evolution process and capture the process as video data using a high frame camera, where a dataset of evolving cracks is created consisting of rock crack images that are manually labeled in pixel-level granularity. We propose a two-stage method to detect cracks in video data: the first stage employs Convolution Neural Network (CNN) based deep learning method to obtain preliminary results for each image frame while the second stage relies on novel variant Bayesian Inference to further refine the detection results. Specifically, in the first stage, a variant of U-Net model (denoted as CrackUNet) is developed to obtain intermediate classifications (crack or non-crack) that can better combine with other processing techniques for further improvement. Then in the second stage, a novel Spatial-Temporal Bayesian Inference (STBI) method is developed to further improve detection accuracy by taking advantages of the spatial and temporal correlations of the evolving cracks in video data. Experimental results show that the proposed method outperforms all the baselines.