Human Activity Recognition (HAR) has earned a lot of importance in recent years due to its applications in various domains including smart healthcare, entertainment, surveillance applications and so on. Due to easy portability and privacy, inertial sensing based HAR has gained potential research interests. The accelerometer and gyroscope sensors detect the body acceleration and angular acceleration respectively which are essentially time series signals. However, 1D time-series signal patterns are found to be insufficient for recognition of some of the daily activities. Existing HAR works are mostly focused on the recognition of a given bunch of activities rather than finding out what kind of data dimensions and classifiers work best to identify different kind of activities. In this paper, we propose a two-phase deep learning classification approach for activity recognition that handles the problem hierarchically. We have shown that signals for different groups of activities need different preprocessing measures in order to extract effective high dimensional features from them in a later phase. CNNs with different dimensions have been utilized accordingly for automatic extraction of robust features. A combination of temporal pattern analysis and spatial pattern analysis is applied through the proposed approach. SVM has been used to divide the bunch of activities in two classes, static and dynamic. The proposed framework achieves overall accuracy of 95.72% for all activities of the UCI HAR dataset when the training data and the test data are collected from different user subsets. Thus, the results indicate that this approach can be applied to classify data from new set of users as well.