查看文章

One of the most important tasks in a visual perception system for automotive navigation and Navigable Path Detection (NPD) is the perception of the environment and detection of obstacles. This task is a challenge still to be overcome for navigation on heavily damaged and unpaved roads. To address this, we applied methods previously developed by us, in particular multi-resolution methods and successive training methods, and new, state-of-the-art, multi-resolution methods using only the TAS500 dataset in the context of the DAGM GCPR 2021 Outdoor Semantic Segmentation Challenge. As our multiresolution approach, we investigated a classical U-Net architecture using the one cycle training policy and an adaptation of a successive growing resolution training strategy. As a second solution, we employed the HRNet model, which has an architecture that naturally learns multiple resolutions during the training, in two versions: the original implementation and a refined Semantic Segmentation version, called SemTorch, which also employs the one cycle training policy. The results show that SemTorch HRNets have the potential to be architectures-of-choice for embedded navigable path detection as they achieved 0.659 mIoU in the test set (third place in the challenge) and 15.74 FPS in a TensorRT engine on an NVIDIA JETSON TX2.