The field of human activity recognition (HAR) using machine learning approaches has gained a lot of interest in the research community due to its empowerment of automation and autonomous systems in industries and homes with respect to the given context and due to the increasing number of smart wearable devices. However, it is challenging to achieve a considerable accuracy for recognizing actions with diverse set of wearable devices due to their variance in feature spaces, sampling rate, units, sensor modalities and so forth. Furthermore, collecting annotated data has always been a serious issue in the machine learning community. Domain adaptation is a field that helps to cope with the issue by training on the source domain and labeling the samples in the target domain, however, due to the aforementioned variances (heterogeneity) in wearable sensor data, the action recognition accuracy remains on the lower side. Existing studies try to make the target domain feature space compliant with the source domain to improve the results, but it assumes that the system has a prior knowledge of the feature space of the target domain, which does not reflect real-world implication. In this regard, we propose stacked autoencoder and meta-learning based heterogeneous domain adaptation (SAM- HDD) network. The stacked autoencoder part is trained on the source domain feature space to extract the latent representation and train the employed classifiers, accordingly. The classification probabilities from the classifiers are trained with meta-learner to further improve the recognition performance. The data from tar- get domain undergoes the encoding layers of the trained stacked autoencoders to extract the latent representations, followed by the classification of label from the trained classifiers and meta- learner. The results show that the proposed approach is efficient in terms of accuracy score and achieves best results among the existing works, respectively.