Crab larvae total estimation is of great significance in crab breeding; however, the existing total estimation methods are complicated and unpractical. Therefore, an estimation method based on computer vision and non-parametric estimation was presented in this paper. Firstly, the videos were captured at different layers of the breeding pond, the foreground image was obtained by using image processing, and the number of crabs in the video was calculated by watershed algorithm and contour extraction to be used as the sample data. Then the probability density function was obtained by kernel density estimation. Finally, combining with the function and sample data, the total number of crabs was estimated. The results showed that the average correct rate of the method was 82.14% when the density of crab was between 100 and 160 liters per liter for small crab ponds with a capacity of 1000L. The study not only solved the complicated operation and crabs transferring problems in capturing video, but also solved the interference of background impurities in the image processing. The method could also be extended to the estimation of many other species such as shrimps and fish fries, and has good versatility and feasibility.