FARMLAND OBSTACLE RECOGNITION BASED ON IMPROVED FASTER R-CNN

For the accurate detection of obstacles in complex farmland environments, ResNet50 is adopted as the backbone feature extraction network, feature pyramid network (FPN) is utilized to enhance the multi-scale feature fusion capability, and the region of interest alignment (ROI Align) strategy is introduced to improve the candidate box localization precision. The experimental results show that the precision, recall, and mean accuracy (mAP) of the improved model are 91.6%, 89.7%, and 93.8%, respectively, which are improved by 2.7, 2.3, and 3.1 percentage points compared with the original base network, and provide a technical reference for navigation and obstacle avoidance of unmanned agricultural machinery.