People Counting Application with Crowded Scenarios: A Case Study with TV Boxes as Edge Devices

Counting people in various urban spaces using artificial intelligence enables a wide range of smart city applications, enhancing governance and improving citizens' quality of life. However, the rapid expansion of edge computing for these applications raises concerns about the growing volume of electronic waste. To address this challenge, our previous work demonstrated the feasibility of repurposing confiscated illegal TV boxes as Internet of Things (IoT) edge devices for machine learning applications, specifically for people counting using images captured by cameras. Despite promising results, experiments in crowded scenarios revealed a high Mean Absolute Error (MAE). In this work, we propose a patching technique applied to YOLOv8 models to mitigate this limitation. By employing this technique, we successfully reduced the MAE from 8.77 to 3.77 using the nano version of YOLOv8, converted to TensorFlow Lite, on a custom dataset collected at the entrance of a university restaurant. This work presents an effective solution for resource-constrained devices and promotes a sustainable approach to repurposing hardware that would otherwise contribute to electronic waste.