The traditional principal curve algorithm is widely used in many fields, but it is ineffective in extracting the principal curves for complex data. To solve the kind of the problem, one of most effective ways is to combine the granular computing with the principal curve algorithm. Therefore, a new multi-granularity principal curve extraction algorithm for complex data based on granular computing is proposed. Firstly, we use the spectral clustering algorithm based on t-nearest neighbor (TNN) to granulate the data and propose the inflexion point estimation to automatically determine the number of granules. Then the local principal curve extraction for each granule is carried out by using soft K-segments principal curve algorithm and optimized by removing the false edges. Finally, a local-to-global strategy is adopted to extract the multi-granularity principal curves to optimize overfitting curves and a principal curve which can describe the original data distribution pattern can be obtained. Experimental results demonstrate the excellent feasibility of the proposed principal curve extraction algorithm.