Abstract: In this paper,a novel scale-invariant local feature extraction method was presented for non-rigid 3D model,called LMP-HKS (local multilevel pattern based heat kernel signatures). Firstly,a coding method was designed based on the local multilevel pattern to describe 1D (one dimension) signal suitably. Then the feature histogram was computed with local multilevel pattern to describe the logarithm difference of the heat kernel signatures. Analyses results show that,the proposed method can not only retain the excellent characteristics of HKS feature,such as isometric invariance,completeness and stability et al,but also keep the local feature scale invariant. Compared with SI-HKS feature that is also scale invariant,the LMP-HKS feature shows better descriptive ability for local shape structure of the 3D non-rigid model. The effectiveness of the proposed local feature has been validated by extensive non-rigid 3D model retrieval experiments.