YOLO-Based Obstacle Detection Under Few-Shot Learning

The obstacle detection industry has been developing rapidly in recent years. Obstacle detection technology has grown by leaps and bounds. Few-shot obstacle detection has a profound significance, which can help us to solve the problem of small samples under high speed conditions in target detection. Due to the limitation of the shooting environment, the sample size about the obstacles is small today. The goal of small sample learning is not for the machine to recognize images in the training set and generalize to the test set; the goal of small sample learning is for the machine to learn to learn on its own. The purpose of learning is for the model to understand the similarities and differences between things and learn to distinguish between them. In this paper, we help with this process by adding a parameter reweighting module. The models will fall into overfitting when a small number of samples are available. In this paper, we propose a small-sample detection method based on improved YOLO-v8. For the problem of small sample size, we improve the generalization ability of the model by adding an attention mechanism at the detection header, and for the problem of model overfitting, we update the model parameters by adding a feature reweighting module to improve the model effect. Finally, the proposed method is validated based on the track area dataset, and the experimental results show that the model can realize the accurate recognition of small sample obstacles.