Human activities can be grouped into basic activities, complex activities and postural transitions within or between basic activities. In the literature, fewer works are focusing on postural transition recognition. Postural transitions, i.e., stand-to-sit and sit-to-stand, etc., are regular functional tasks indicative of muscle balance and power performance. Hence, examining postural transitions delivers an objective tool for a human mobility assessment. In this paper, a smartphone-based human activity recognition architecture with the utilization of multi-head temporal convolutional operators is proposed. In this work, a custom-tailored learning architecture for each sensor type, i.e., accelerometer and gyroscope sensors in this case, is developed. Feature learning is enhanced by deliberating on a single sensor instead of on all at once. This allows better data characterization of the multichannel inertial data. Empirical results reveal the superiority of the proposed model to the existing machine and deep learning models, with an accuracy of 98.35%.