In order to effectively improve the activity classification efficiency in body sensor networks, a maximum likelihood sparse representation algorithm based on K-SVD is proposed in this study. Firstly, all of activity pattern training samples are grouped according their classes to be trained, respectively. The mutual interference among different groups in the process of training can be avoided and sub-dictionaries for every class can be obtained. Then, these sub-dictionaries are used to construct an over-complete dictionary. And the dictionary is able to sparsely represent the testing samples precisely. The sparse representation coefficients are precisely approximated by maximum likelihood sparse model and the recognition result of testing samples are determined by the coefficients. The experimental results show that the proposed algorithm is able to obtain the optimal dictionary and the method based on maximum sparse representation can precisely estimate the representation error of testing activity samples. The accuracy of the proposed algorithm is obviously better than some conventional sparse-representation-based activity recognition algorithms. The proposed algorithm is able to effectively improve the activity pattern classification efficiency in body sensor networks.