Efficient Device Free Crowd Counting by Leveraging Transfer Learning

Recently, wireless sensing is gaining immense attention in the Internet of Things (IoT) for crowd counting and occupancy detection. As wireless signals propagate, they tend to scatter and reflect in various directions depending on the number of people in the indoor environment. The combined effect of these variations on wireless signals is characterized by the channel state information (CSI), which can be further exploited to identify the presence of people. State-of-the-art CSI-based supervised crowd counting systems are vulnerable to temporal and environmental dynamics in practical scenarios as their performance degrades with fluctuations in the indoor environments due to multipath fading. Inspired by the breakthroughs of transfer learning and advancement in edge computing, we have leveraged in this work the concept of transfer learning to minimize this problem via exploiting the trained model from the source environment for other indoor environments to perform device-free crowd counting (CrossCount) at the target rooms. Our results show that this technique can combat the dynamics of the environment and achieves 4.7% better accuracy with 40% reduction in training time as compared to conventional convolutional neural networks. In essence, our results imply the future possibility of harnessing crowdsourced CSI data collected at different indoor environments to boost the accuracy and efficiency of local crowd counting systems.