As a new initiative by smart city projects that are going viral in worldwide ICT developments, mostly by governments, IoT sensors and their applications have been exploited and adopted proactively nowadays. A useful but relatively low-tech application is counting human presence by using carbon dioxide sensor. Such CO2 sensors are durable and inexpensive, with their compact sizes they could be deployed anywhere for estimating head counts ubiquitously. In this paper, a case study of applying CO2 sensors in public buses is investigated. Counting passengers in public buses or public transport in general has great economics advantages. However, a few technical challenges include but not limited to the mobility of the bus, the dynamic air flows, and factors such as windows were open, ventilation and even urban pollution etc, would affect the accuracy of occupancy counting. Hardly there would be a simple linear mapping between the number of people in a bus and the measurement of CO2 level. Hence, non-linear machine learning tool is used for inferring the non-linear relation between the two, with the consideration of the mentioned influential factors. Empirical data are collected from experiments conducted in several different buses over different times. The results can point to a promising conclusion that satisfactory accuracy could be achieved.