In this paper, we present a novel approach based on gait energy image (GEI) and co-evolutionary genetic programming (CGP) for human activity classification. Specifically, Hu?s moment and normalized histogram bins are extracted from the original GEIs as input features. CGP is employed to reduce the feature dimensionality and learn the classifiers. The strategy of majority voting is applied to the CGP to improve the overall performance in consideration of the diversification of genetic programming. This learningbased approach improves the classification accuracy by approximately 7 percent in comparison to the traditional classifiers.