Recently, lower limb activity recognition (LLAR) based on surface electromyography (sEMG) signal has attracted increasing attention, mainly due to its applications in the control of robots and prosthetics, medical rehabilitation, etc. Traditional machine learning-based LLAR methods rely on expert experience for feature extraction. In addition, the noise interference and class-imbalanced problem can also affect the recognition effect. Aiming at these problems, a LLAR framework based on sEMG data augmentation (DA) and enhanced capsule network (ECN) is proposed in this paper. Firstly, a hybrid denoising technique combining variational mode decomposition and non-local means estimation is designed to effectively filter out noise components mixed in the sEMG. Then, K-Means synthetic minority oversampling technique is utilized to synthesize new samples for minority classes, thereby overcoming the influence of class imbalance on recognition reliability. Finally, an ECN model is constructed to implement end-to-end LLAR, in which an efficient channel attention module is embedded to mine sensitive features, thus further improving the feature learning ability of the classifier. Experimental results indicate that the proposed framework is applicable to multiple types of individuals, including healthy subjects, patients with knee abnormalities, and patients with stroke, providing more satisfactory recognition performance and robustness than state-of-the-art methods..