Sensor-based human activity data (HAD) collection, used for engineering human activity recognition (HAR) systems, is a cumbersome process. Its success depends on two factors: correct activity labels and correct sensor activation sequence. This is very hard to achieve, as is seen in many previous studies. Therefore, a system for automatically observing HAD collection is introduced, making it a lot easier and less noisy. At the heart of the system sits a novel data mining tool that uses Web activity data to achieve the above goal. The effectiveness of the proposed system is shown using a set of two experiments.