TL;DR
This paper tackles the challenge of accurately estimating pedestrian flow rates in crowded events by fusing data from ubiquitous Wi-Fi sensors with a limited number of automated counting systems. They found that Recurrent Neural Networks (RNNs), particularly LSTMs, combined with contextual information about the overall crowd state, significantly improve flow rate estimation compared to Wi-Fi data alone.
What Problem Does This Paper Address?
Crowd monitoring systems (CMSs) are crucial for managing large gatherings, but relying on a single sensor type limits the insights they can provide. While Wi-Fi sensors are widespread and can track individual devices, they offer a sparse view of crowd density. Automated counting systems (like cameras) provide more direct flow rate data but are expensive to deploy everywhere. The problem is how to get a comprehensive understanding of crowd dynamics, specifically pedestrian flow rate, without incurring the prohibitive cost of equipping every location with expensive sensors. This research aims to bridge this gap by using data fusion to enhance the capabilities of a Wi-Fi-centric CMS with a few strategically placed automated counters.
Methods
The core of this research lies in exploring various data fusion models to estimate pedestrian flow rate. They tested several approaches:
- Indirect Regression Models: These models first estimate a global "multiplication factor" relating Wi-Fi counts to flow rate, then use this factor to predict flow. They explored linear, quadratic, and logarithmic models for this factor.
- Direct (Multiple) Regression Models: These models directly relate flow rate to Wi-Fi counts, incorporating the variance in the data more directly. They tested linear, quadratic, and exponential forms.
- Shallow Neural Networks (NN): A single-layer neural network was used, capable of learning non-linear relationships.
- Autoregressive Moving Average (ARMAX) Model: This model explicitly incorporates past time-series data to predict future values.
- Recurrent Neural Network (RNN-LSTM): This advanced neural network architecture is designed to handle sequential data and maintain memory of past states, making it well-suited for time-series prediction.
Crucially, they investigated two scenarios for each model type: one using only local Wi-Fi counts, and another that also incorporates "contextual variables." These contextual variables are aggregate statistics from the entire sensor network, such as the total Wi-Fi count across all sensors, average flow rate, and changes in flow rate. This allows them to assess the benefit of spatial correlation and global crowd state information.
Datasets
The study utilized data collected from a real-world, four-day music and motorsport event, the TT Festival in Assen, Netherlands, in 2018. The data comprised:
- Automated Counting System Data: Provided ground truth flow rate measurements from five camera-based counting systems (C1-C5) at key locations.
- Wi-Fi Sensor Data: Collected MAC addresses from Wi-Fi-enabled devices, which were processed to derive Wi-Fi counts (a proxy for pedestrian density).
The data was split into training (Wednesday and Friday) and validation (Thursday and Saturday) sets. The authors note that the data was collected during specific evening hours (6 PM to 11 PM) to mitigate measurement errors associated with changing light conditions at night.
Key Insights and Ideas
- Wi-Fi alone is insufficient for accurate flow rate: The scatter plots clearly show that Wi-Fi counts alone don't directly correlate perfectly with flow rates, especially during peak times or in areas with mixed pedestrian uses (e.g., standing to watch performances vs. moving through).
- Context is king: Incorporating aggregate data from the entire sensor network (like total Wi-Fi counts and average flow rates) significantly improved the performance of most models, highlighting the spatial correlation of crowd movement.
- RNNs excel at capturing dynamics: The RNN-LSTM model, with its ability to learn temporal dependencies, emerged as the best performer, suggesting that the dynamic nature of crowd flow is critical.
- Population characteristics matter: The models struggled with high flow rates and showed sensitivity to changes in the visitor population (e.g., smartphone usage patterns), indicating a need for models that are robust to these variations or can explicitly account for them.
Key Results
The study identified the RNN-LSTM model with one hidden node, incorporating contextual variables (specifically, average total flow rate and the average ratio between Wi-Fi counts and flow rate) as the best performing model.
- This model achieved the highest adjusted coefficient of determination ($R^2_{adj}$) of 0.4306 and the lowest RMSE of 11.77 on the validation dataset.
- It outperformed simpler models like linear regression and shallow neural networks, especially in capturing the general trend of flow rates, particularly at lower flow rates which occur most frequently.
- While the best models could predict the general trend, they struggled to accurately capture peak flow rates and sharp increases, often underestimating values above 50 pedestrians per minute.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Sensor Location Specificity: The current models are not specialized per sensor location, and the influence of specific sensor location characteristics (e.g., being a thoroughfare vs. an area for standing) on the Wi-Fi count-to-flow rate ratio needs further investigation.
- Model Structure Optimization: The study explored a limited set of neural network structures; deeper learning architectures might offer further improvements.
- Population Representativeness: The model was trained on a specific festival population. Its applicability to other demographics or event types (e.g., concerts, theme parks) needs to be tested, as population characteristics can influence smartphone usage and thus Wi-Fi trace data.
- Wi-Fi Data Quality: The increasing use of rotating MAC addresses and evolving privacy protocols can make Wi-Fi data noisier and harder to process. Future work should focus on data fusion methods that are more robust to such data quality issues.
- Mixed-Use Spaces: The presence of areas where pedestrians engage in multiple activities (traveling, watching performances) complicates the direct translation of Wi-Fi counts to flow rates. Further research is needed to disentangle these behaviors.
Summary updated