Sensor-based human activity recognition (HAR) plays a major role in healthcare and security applications. The significance of this study is to understand the state-of-art of techniques for recognizing the activities of humans based on physiological signals acquired by a body-worn sensors. Accurate recognition of activities with just signals from the wearable sensor is a difficult task due to the inherent complexity of physical activities. Even though sensor-based HAR has been accomplished by using various algorithms exploiting machine and deep learning techniques, just a handful of researchers have made an extensive study on the contribution of various parameters on the accuracy of recognition of human activity. The main focus of this study is to perform a comparative evaluation of the state-of-the-art algorithms based on machine and deep learning techniques that have been proposed for HAR. Principal component analysis (PCA) and t-distributed stochastic neighbour embedding (t-SNE) methods are employed for data dimensional reduction and visualization. Machine learning algorithms like random forest (RF), kernel-based support vector machine (SVM), and deep learning algorithms such as convolutional neural networks (CNN) have been applied on the University of California, Irvine (UCI) HAR dataset. A comprehensive study has been conducted to understand the impact of changing parameters like pooling, activation functions, number of dense layers, and dropout percentage of CNN on the accuracy of recognition. The proposed work employed the swish activation function in the dense layer which was recently proposed by Google. The output layer of the last dense network employs the softmax function to classify human activities. The proposed CNN architecture (with max-pooling layer, the swish activation function in one dense layer, and softmax function at the output layer) achieved results of training and validation accuracy as 99.58% and 92.57% respectively for HAR.