Robust pixel-wise detection of road obstacles by integrating composite and real images

Safe navigation systems for vehicles use semantic segmentation and other object identification technologies. However, these neural networks are trained on a closed set of known classes, and because of overconfidence in their predictions, it is difficult for them to identify unseen objects. In this study, we propose a framework to identify road obstacles in the drivable area of a driving scene. We built our model on the Synboost framework, integrating a maximized entropy module to generate high-quality priors and image resynthesis as an attention mechanism for a spatial-dissimilarity network. The architecture of our dissimilarity network uses prior images to create a single-channel modulation map at each level of the decoder to weigh features. Moreover, we used a mixture of real and synthetic anomalies during training and fine-tuning. For evaluation and testing, we used only real anomalies. Our experiments demonstrated the importance of using multiple sources of anomalies during training for robust pixel-wise anomaly detection, as all models consistently yielded better results when trained on a combination of real and synthetic images. Additionally, the experimental results demonstrated that, at a 95% true positive rate, our proposed model outperforms previous models in average precision and false positive rate on two test sets.