The fascinating innovation that is Human Activity Recognition (HAR), specifically the sensor-based type, has unlocked a sea of applications in diverse spheres of today’s tech-powered life. This work delves into the performance evaluation findings of four Machine Learning classifiers: Random Forests, SVM, Neural Network and kNN and four Deep Learning frameworks: CNN, LSTM, GRU and CNN + LSTM, achieved on the sensor-based UCI-HAR dataset with the aid of necessary software environments. To ascertain the ideal model for sensor-based HAR, the required training and testing experiments for both ML and DL approaches are proceeded using a two-fold strategy. The initial phase involved completing basic model implementations of the aforementioned techniques which helped form an estimate of the preliminary efficiency levels. Thereafter, we engaged hyperparameter tuning algorithms to elevate the recognition performance of each ML and DL model even more. Confusion matrices so obtained before and after tuning broadened the perspective in regard to the extent to which the activity samples are correctly classified plus those activities that see the most number of confusions. Additionally, insightful observations about hyperparameter tuning’s impact in enhancing overall performance results are garnered. Above all, accuracy comparison results revealed that SVM is the best-performing model for both before and after optimization contexts with 96.40% and 96.54% accuracy respectively. Both versions of SVM thus not only outperform the rest of the ML models but even the selected DL-based models.