Heatmap Regression-Based Context-Aware Learning for Tillage Boundary Detection

In agricultural automation, autonomous tractors play a crucial role in enhancing farming efficiency, particularly through the navigation and identification of tillage boundaries. Traditional approaches rely on machine vision techniques to identify paths by distinguishing between tilled and untilled soil areas. Although recent advancements in convolutional neural networks (CNNs) have shown promising results in agricultural automation, they still face challenges in fully capturing the global context of tillage boundaries. This limitation stems mainly from CNNs’ small receptive fields, which often limit the network’s capacity to capture broader contextual information in agricultural landscapes, potentially causing inaccuracies in boundary detection. These methods rely significantly on local feature analysis and necessitate complex, computationally intensive heuristic post-processing to enhance detected tillage lines, thus limiting their real-time application efficacy. We propose a line-context-aware learning method that combines a heatmap regression with a transformer to more effectively learn and extract global contextual features. The proposed end-to-end method streamlines detection, enhancing real-time agricultural applications and improving the accuracy and reliability of autonomous tractor navigation and operation. The proposed method was evaluated on a custom dataset, demonstrating competitive performance in accurately detecting tillage boundaries and proving its capability to handle the intricate details and variations present in agricultural landscapes.