Definition of object tracking is identifying the moving targets from complex background, and it should track the target accurately and continuously. In occlusion, deformation, complex background conditions robust tracking target is still a challenging problem to be solved. A novel online object tracking algorithm is proposed for the occlusion and deformation with sparse prototypes, which exploits local linear embedding (LLE) algorithm with sparse representation scheme for learning effective appearance model. LLE is a classic manifold learning algorithm. In the algorithm, the neighbor points weight of each point remains unchanged in translation, rotation, scale changes. Thus, it can be used to extract the essential characteristics of target and find the inherent law of data. Firstly, the algorithm uses the local linear embedding algorithm to extract low dimensional characteristic. Then the sparse prototype is composed of the base vector which is extracted from the low dimensional characteristic and trivial templates. It can be used to update templates. This algorithm maintains the advantages of the original sparse tracking method to occlusion, and has a good robust tracking effect of deformation object. The experimental results show that the proposed tracking algorithm is better than the other seven commonly used algorithms in the nine video sequences.