Real-Time Obstacle Avoidance on Autonomous Lawn Mowers using Deep Learning

Autonomous robotics are increasingly integral to modern technological advancements, offering the potential to transform industries by enabling machines to perceive and navigate their environments independently. Vision-based obstacle avoidance is a critical component of these systems, ensuring safe and efficient operation in dynamic settings. However, deploying sophisticated deep learning models on resource-limited devices, such as autonomous lawn mowers, requires innovative approaches to balance computational efficiency with predictive accuracy. This thesis presents a practical integration of deep learning models on autonomous lawn mowers, addressing the vision-based obstacle avoidance problem. The primary focus is enhancing the robustness and efficiency of obstacle detection models while leveraging their predictions to implement avoidance strategies. To tackle these challenges a representative dataset is collected, covering different environmental conditions. Key techniques, such as pruning and quantization, are applied to reduce models size, making them feasible for deployment on resource-constrained devices. Additionally, a practical obstacle avoidance pipeline is designed, which not only relies on model predictions but is also enhanced by an innovative, simple but effective approach that incorporates color information from the input images. The proposed method provides high efficiency and accuracy, demonstrating strong performance in real-world scenarios. The quantitative analysis of the results reveals a critical trade-off between efficiency and accuracy, providing valuable insights into optimizing deep learning models for real-time applications in autonomous lawn mowers.