Real-time Crowd Measurement in GPS Denied Open Areas
Introduction
In the Covid era, crowd measurement is considered a great challenge as it is required almost everywhere, indoor and outdoor. We can see many systems have been introduced for indoor places (such as supermarkets, classrooms, workplaces) to count visitors by providing hardware and software-based solutions. These solutions are using a set of different technologies, including Infrared, Thermal Imaging, Video-Based, and others.
Estimating pedestrians’ locations is not an easy task, especially for open areas with 1) poor Global Positioning System (GPS) signal reception and/or 2) lack of infrastructure to install expensive solutions such as video-based systems.
Street furniture and crowd measurement
Street furniture assets such as garbage bins, seats, and bus shelters exist where pedestrians exist; accordingly, street furniture provides excellent coverage across open areas such as parks, streets, homes, universities. This makes them a great candidate for measuring the crowd.
In this article, we will show a low-cost solution driven from my Ph.D. thesis based on street furniture[1] by extending the street furniture role to provide a real-time localization system.
Can we have a low-power, low-cost and low-maintenance sensor to enable street furniture to measure the crowd in real-time?
To answer this question, we need to find a technology that is reliable and doesn’t require a massive investment in terms of infrastructure.
Several relevant works have proposed different methods based on crowed video analysis, Geotagged Social Media (GTM), Radio Frequency Identification (RFID), Public Transportation Smart Card (PTSC), GPS, Cellular Service Records (CSR), WIFI Received Signal Strength (RSS), and Bluetooth Detection Records (BDR) to understand human mobility.
RSS-based methods exploit the Wifi-enabled devices standard to measure crowd behaviour in which mobile devices frequently send Wifi probe requests to the surrounding Access Points (APs) to accelerate the Wi-Fi connection process. These probe requests contain Media Access Control (MAC) addresses, and RSSs are recorded by the receiver. They outperform other methods because 1) unlike GPS, they provide fine-grained spatial and temporal data given that there are suitable amount of APs covering the area, 2) they provide high precision localization,3) they can measure crowd behaviour for different types of crowds in different locations and modes such as transportation and 4) even if users disabled WIFI on their devices, mobile devices keep sending prob requests[2,3].
Using a simple wireless sensor node (Arduino and ESP-13 module), we can write a program to configure the ESP-13 module to work as a soft access point to sniff probe requests. Moreover, the sensor communicates data inside the probe requests (signal strength and MAC address) with the centralized server for future analysis and location prediction.
For better power management and prolonging battery life, the server can configure the devices in terms of the frequency of sending data and/or active/sleep mode. For example, the server may change the mode of the device from active to sleep mode.
Real-time Localisation Approach based on a four-layer deep neural network
Before building a prediction model, we need to generate a labelled dataset using some users' devices. To get a labelled dataset, it is required to grid the area into a two-dimensional array. Then, some users use a developed mobile app to send their location at a specific time, and the server uses MAC address and time to label the dataset. We applied the following rules to clean the data. 1- Under certain circumstances (e.g., limited/no Internet connection), APs could not send data, which affects the quality of the prediction model accuracy. Therefore, the records not containing MIN_NO_RSSs greater than -100 will be excluded.
2- We did not use the actual RSSs values but instead normalised the values using Zscore and removing any outliers.
Finally, a four-layer deep neural network classifier. The classifier is used twice in parallel to predict x and y. The input layer has the no_AP nodes, which is the number of APs that contribute to fingerprint records. Then, three hidden layers with a power of 2 nodes. Finally, the output layer has only one node. To improve model generalization and reduce overfitting, we used the dropout regularisation method [5] in which, during training, in each layer, a random number of outputs are dropped out.
Experiments
Experiments conducted on the Murdoch University campus with results showing that the proposal improves the lifetime of wireless sensor nodes by up to 412% compared to existing architectures [4]. Moreover, network traffic is improved by up to 38% without affecting data freshness. Finally, storage space for the database at the server is reduced by up to 99%. Based on the prediction approach, accuracy reached 98%.
Conclusion
There is an urgent need for real-time crowd measurement for open areas that don’t have GPS and/or well-established infrastructure. WIFI RSS-based methods outperform other methods when it comes to the scalability, cost, availability, and precision of data. It explained how smart cities can utilize existing street furniture to provide a better understanding of pedestrians’ movements and flow in real-time using RSS.
An efficient WSN architecture for smart street furniture was presented to reduce power consumption, storage, and network traffic without affecting data freshness. A localization approach that uses data cleaning, feature engineering, and a four-layer deep neural network classifier was provided. The approach achieves 98% accuracy.
References
[1] Ali, Mohamed (2020) Real-time localisation system for GPS denied open areas using smart street furniture. Ph.D. thesis, Murdoch University.https://researchrepository.murdoch.edu.au/id/eprint/59730/
[2] Freudiger, Julien. “How talkative is your mobile device? An experimental study of Wi-Fi probe requests.” Proceedings of the 8th ACM Conference on Security & Privacy in Wireless and Mobile Networks. 2015.
[3] https://hal.inria.fr/hal-01575519v2/document
[4] Nassar, Mohamed A., et al. “Adaptive low-power Wireless Sensor Network architecture for smart street furniture-based crowd and environmental measurements.” 2019 IEEE 20th International Symposium on” A World of Wireless, Mobile and Multimedia Networks”(WoWMoM). IEEE, 2019.
[5] Pham, V., et al. Dropout improves recurrent neural networks for handwriting recognition. in 2014 14th International Conference on Frontiers in Handwriting Recognition. 2014. IEEE.
