TreeHelper: A Wood Transport Authorization and Monitoring System

TL;DR

TreeHelper is an IoT system that uses a web platform and edge AI to monitor and authorize wood transport, helping authorities detect unauthorized logging more efficiently.

Contribution

TreeHelper introduces an edge AI device with a custom-trained YOLOv11n model and a web platform for real-time wood transport monitoring and authorization.

Findings

Edge TPU acceleration improves throughput and reduces inference latency for real-time monitoring.

The YOLOv11n model achieves high validation performance with precision and recall above 0.94.

Unauthorized trucks trigger alerts via SMS with license plate and GPS data for rapid authority response.

Abstract

This paper proposes TreeHelper, an IoT solution that aims to improve authorization and monitoring practices, in order to help authorities act faster and save essential elements of the environment. It is composed of two important parts: a web platform and an edge AI device placed on the routes of tree logging trucks. The web platform is built using Spring Boot for the backend, React for the frontend and PostgreSQL as the database. It allows transporters to request wood transport authorizations in a straight-forward manner, while giving authorities the chance to review and decide upon these requests. The smart monitoring device consists of a Raspberry Pi for processing, a camera for capturing live video, a Coral USB Accelerator in order to accelerate model inference and a SIM7600 4G HAT for communication and GPS data acquisition. The model used is YOLOv11n and it is trained on a custom dataset of tree logging truck images. Model inference is run on the frames of the live camera feed and, if a truck is detected, the frame is sent to a cloud ALPR service in order to extract the license plate number. Then, using the 4G connection, the license plate number is sent to the backend and a check for an associated authorization is performed. If nothing is found, the authorities are alerted through an SMS message containing the license plate number and the GPS coordinates, so they can act accordingly. Edge TPU acceleration approximately doubles TreeHelper’s throughput (from around 5 FPS average to above 10 FPS) and halves its mean inference latency (from around 200 ms average to under 100 ms) compared with CPU-only execution. It also improves p95 latency and lowers CPU temperature. The YOLOv11n model, trained on 1752 images, delivers high validation performance (precision = 0.948; recall = 0.944; strong mAP: mAP50 = 0.967; mAP50-95 = 0.668), allowing for real-time monitoring.

Figures