Abstract
Urban data mining can be identified as a highly potential area that can enhance the smart city services toward better sustainable development especially in the urban residential activity tracking. While existing human activity tracking systems have demonstrated the capability to unveil the hidden aspects of citizens' behavior, they often come with a high implementation cost and require a large communication bandwidth. In this article, we study the implementation of low-cost analog sound sensors to detect outdoor activities and estimate the raining period in an urban residential area. The analog sound sensors are transmitted to the cloud every 5 min in histogram format, which consists of sound data sampled every 100 ms (10 Hz). We then use wavelet transformation and principal component analysis to generate a more robust and consistent feature set from the histogram. After that, we performed unsupervised clustering and attempt to understand the individual characteristics of each cluster to identify outdoor residential activities. In addition, on-site validation has been conducted to show the effectiveness of our approach.