Low-latency perception in off-road dynamical low visibility environments

This work proposes a perception system for autonomous vehicles and advanced driver assistance specialized on unpaved roads off-road environments. In this research, the authors have investigated behavior of Deep Learning algorithms applied to semantic segmentation environments under differents adverse conditions visibility. Almost 12,000 images different were collected labeled. It was assembled an proving ground exclusively its development. The proposed dataset also contains many situations such as rain, dust, low light. To develop system, we used convolutional neural networks trained segment obstacles areas where car can pass through. We developed Configurable Modular Segmentation Network (CMSNet) framework help create architectures arrangements test them dataset. Besides, ported some CMSNet configurations by removing fusing layers using TensorRT, C++, CUDA achieve embedded real-time inference allow field tests. main contributions are: new off-roads containing night, dust; framework; investigation regarding feasibility applying deep learning detect region vehicle through when there is no clear boundary track; study how our behave in severity levels visibility impairment; evaluation tests carried out with inference.