This paper proposes PedestrianCrowd-Net, a deep learning framework that uses commercial cellular signals (LTE/5G) to differentiate pedestrian activities in dense urban environments without requiring visual sensors or device cooperation. The approach exploits fine-grained channel state information (CSI) and reference signal received power (RSRP) variations to capture motion signatures, using a multi-stream temporal convolutional network with attention mechanisms. The method addresses challenges including inter-cell interference, multipath propagation, and activity differentiation in mixed crowds. Experimental results demonstrate classification accuracies exceeding 85% for individual activities and 78% for mixed crowd scenarios, establishing a foundation for scalable, privacy-preserving crowd monitoring using existing cellular infrastructure.
Key findings
Commercial cellular signals (LTE/5G) can achieve over 85% accuracy for individual pedestrian activity classification and 78% for mixed crowd scenarios when processed with appropriate deep learning architectures.
A multi-stream temporal convolutional network with attention mechanisms effectively processes cellular signal features across multiple time scales to capture both individual and collective activity patterns.
Signal preprocessing techniques can mitigate inter-cell interference (ICI) and multipath propagation effects in dense urban environments while preserving activity-sensitive features.
The proposed approach enables privacy-preserving crowd monitoring without requiring cameras, wearable devices, or active user participation, leveraging existing cellular infrastructure.
Limitations & open questions
The study relies on preliminary analysis and proposed experimental designs rather than completed large-scale real-world deployments.
Commercial cellular systems have lower temporal resolution compared to specialized radar or custom WiFi implementations, potentially limiting fine-grained motion detection.
Performance may vary across different urban environments due to varying base station density and multipath conditions.