Sensor Fusion for Public Space Utilization Monitoring in a Smart City (pdf) is simply the best read for IoT product designers, developers and implementers. It steps through designing a system to measure space utilization in a city — the trade-offs made in sensor selection and calibration, power source selection, network design, data cleaning and normalization, and data processing. The methodology can be generalized for designing any IoT network. The paper is nothing less than a perfect case study about how to build an IoT network.
The most interesting aspects of the paper by Billy Pik Lik Lau, Nipun Wijerathne, and Chau Yuen of the Singapore University of Technology and Design and Benny Kai Kiat Ng of Curtin University is how they matched the sensors to acquire the data at the right resolution to estimate space utilization and built a test bed, minimizing a wide range of implementation issues. To measure space utilization, meaning how populated a space is over multiple time intervals, they chose sound and motion sensors and the fusion of the two. The methodology applied in this paper could be adapted to other sensor types.
Sound sensors vs. video sensors
Sound sensors were chosen over video sensors to detect activity. This may seem counter-intuitive, but only because sighted humans' senses are more visually dominant. Cameras are computationally expensive, and the increased processing would require more expensive and more, powerful computers in the network. In a large-scale deployment, this would increase the cost without improving the accuracy of the measurements. More computational capacity would increase the amount of power consumed, exceeding the power supplied by sustainable and low maintenance solar cells. And cloud processing of video data would require significant network bandwidth and storage, increasing cost. Lastly, because of privacy issues, deploying cameras would require permitting, which would pose a problem in deployment in Singapore where the research was done.
The Renewable Wide Area Sensor Network (RWSN) eliminates hardwired power and battery replacement. The renewable power design was not necessary for testing the use of sound and motion sensors because hardwired power or battery power in this limited seven-node network would not have been expensive. Deploying this sensor network at scale presumably was the reason for the renewable choice.
The RWSN uses low-power XBee transmitter modules (IEEE 802.15.4 ) to interconnect the sensors to a Xbee receiver with a Raspberry Pi to backhaul the data to cloud storage. Using Xbee relays, the researchers built a wireless mesh network to increase coverage. The network is powered by solar panels and battery storage. The sizes of both are included, so an IoT system designer at a different lattitude with a different amount of sunlight could adjust the size of both the solar panel and batteries to match local conditions.
An Environmental Monitoring Sensor Node, including a barometer, thermometer, light intensity, resistive rain, ultra-violet (UV) index, humidity, motion, and noise sensors, is included in the network to provide calibration data to remove the effects of environmental conditions, such as rain interfering with sound readings.
Inexpensive Pyroelectric Infrared (PIR) sensors, also known as passive infrared sensors, were used to detect motion. And inexpensive analog sound sensors, essentially microphones, were used to record sound.
The PIR sensors reported many false positives during the daytime, especially during the afternoon. To remove those errors, a calibration module was written to preprocess the data. Onsite ground truth measurement correlated the false positives to bright daylight. The correlation module calculates the probability of false alarms and adjusts for it. The data was then statistically normalized.
Applying machine learning to remove errors
Environmental errors from environmental conditions such as rain were eliminated using an unsupervised machine learning method that allowed the researchers to find the similar patterns of the sound data using clustering. Clustering simply categories similar data sets, such as the sound of rain, which can then be removed from the data. Similarly, background noise can be eliminated.
Normalized and calibrated data from the PIR and sound sensors is fused using an algorithm chosen by the researchers to estimate space utilization throughout the testbed area's seven nodes. Estimation is based on empirical onsite observations and the comparison of the fused data from the seven nodes.
The paper explains the hardware, communications, sensor, and data processing design considerations in building a low-cost accurate IoT system. It also explains the challenges of eliminating false positives from the data captured by the PIR motion sensors and how to characterize the noise signatures from human activity, while eliminating environmental errors such as rain and background noise to accurately estimate utilization of individual points of interest and the testbed as a whole.
IoT networks need the potential to scale to be very large and useful to justify design and development costs against an economic or social return. Design and development required a multidisciplinary team and some specialized skill, especially in sensor engineering and advanced mathematics in the subfields of statistics and machine learning. Mathematics skills were needed to calibrate the data and eliminate background noise in this case. Sensoring engineering skills were needed to acquire data at the right resolution at a low cost. Enterprises, serious about building networks of IoT sensors may need to hire to acquire sensor and mathematics skills.