A Generative Siamese Convolutional Network for Obstacle Classification in Autonomous Driving

The autonomous driving system can reduce traffic accidents caused by human factors, thereby ensuring the personal safety of individuals. The obstacle classification model is a crucial component of the autonomous driving system, assisting vehicles in avoiding potential obstacles and ensuring safe navigation on real roads. However, traditional obstacle classification models have utilized convolutional neural network (CNN) as image feature extraction functions. The obstacle classification models based on CNN exhibit limited generalizability of the learned image features. Additionally, the scarcity and imbalance of labeled data result in poor classification performance, particularly for minority classes of obstacles, in obstacle classification models based on CNN. Therefore, we propose a Generative Siamese Convolutional Network for Obstacle Classification (GSCNOC) in autonomous driving. The GSCNOC model uses a siamese convolutional network to learn images features with better generalizability, thereby enhancing the classification performance on new obstacle images. Features trained in this manner exhibit robust classification and noise resistance while reducing the model's data requirements. Moreover, the GSCNOC model employs a generative adversarial network to augment the quantity of minority class samples, improving the model's training data. We conducted multiple experiments on a real obstacle dataset, validating the superior performance of the GSCNOC model.