An Obstacle Recognition Model Based on Siamese Network With Masked Strategy for Unmanned Aerial Vehicle Obstacle Avoidance

By equipping autonomous aerial vehicles (AAVs) with multiple sensors to gather information about obstacles in their flight environment, we can guide the autonomous and safe flight of AAVs. Existing obstacle avoidance models use cameras to capture environmental images and identify the categories of obstacles within them. However, the environmental images captured by AAVs often contain a significant amount of noise. This noise can interfere with the feature extraction process of the obstacle recognition model, causing it to incorrectly differentiate between various obstacle categories and resulting in decreased classification performance. Additionally, during navigation, AAVs encounter a wide variety of obstacles. Some categories of obstacles are more common, while others are less frequent. This imbalanced distribution of obstacle categories can affect the training process of the obstacle recognition model, leading to lower classification accuracy for certain categories. To address these challenges, we propose an obstacle recognition model based on siamese network with masked strategy (ORSNMS) for AAV obstacle avoidance. The ORSNMS model integrates the advantages of masked autoencoders and DenseNet networks, enabling it to better handle noise and the situation where certain obstacle categories have fewer instances. Specifically, to reduce the interference of noise in the feature extraction process, the ORSNMS model employs a masked strategy to further learn the features of images. By masked part of the data during training, the model can learn more robust image features, improving its performance in noisy environments. Additionally, the ORSNMS model incorporates a DenseNet structure to enhance the training process of categories with fewer samples. By utilizing contrastive loss, the ORSNMS model compares the enhanced features with the original features, minimizing the error between them. The siamese subnetworks in the ORSNMS model share parameters, which not only reduces the number of model parameters but also enhances the model’s generalization capability. The ORSNMS model achieves a Precision of 0.984, Recall of 0.984, and Accuracy of 0.984 on real dataset.