Human Activity Recognition (HAR) is one of the most explored fields of deep learning and sensor technology. HAR is the process of recognising a person’s daily living activities using a set of sensors and an effective pattern generalisation framework. Most of the state-of-the-art deep learning-based HAR models need more efficient model generalisation due to sub-optimal pattern identification by the end classifier. To overcome this issue, this paper presents an effective spatio-temporal feature engineering and classification framework using CNN-LSTM and a Decision tree classifier for a sensor-based HAR system. A custom 1D convolutional and LSTM layer has been developed for extracting local spatial features from raw time-series data and high-level temporal dependencies using a memory unit of recurrent layers. The proposed achieved an average performance accuracy of 98% and 99% with its own and MotionSense dataset, outperforming multiple benchmarks in optimised computational times.