Sensitivity of robot-aided remote object detection in forests under variation of light illumination

Forests degradation and deforestation are increasingly becoming a risk to the world’s ecosystem with major effects on climate change. Mitigating these dangers is tackled through reliable management of monitoring tree species, insect infestations and wildlife behaviour. Although forest rangers can use artificial intelligence and machine learning techniques to analyse forest health through visionary sensing, exploring the accuracy of object detection under low illuminations such as sunsets, clouds or below dense forest canopy is often ignored. In this paper, we have investigated the importance of illumination on detection through a high definition GoPro9 camera as compared to the low-cost RaspberryPi camera. An external sensing platform accommodated by a quadruped robot is developed to carry the hardware, one of the first implementations of autonomous system in forest health monitoring. The compound-scaled object detection, YOLOv5s model pretrained on COCO dataset containing 800,000 instances, used for person detection, is retrained on custom dataset to detect forest health indicators such as burrows and deadwood. The system is tested and evaluated under various lighting conditions to detect objects located at various distances from the vision sensors. This study concludes that YOLOv5s model can detect a person and forest health indicators up to a distance of 10m with accuracy of 67% and 51% respectively at dusk which shows that light exposure has a major effect on detection performance. Furthermore, the quadruped robot carrying the sensing platform managed to successfully shift between different positions while carrying out the detection.