Challenge

Semantic road segmentation is an important task in the context of autonomous driving, which aims to assign each pixel in the visual field of a board camera to an object category like cars, pedestrians, etc. The information from segmented image can be used to identify drivable area, obstacles,  and other  traffic participants.

Solution

We cast the problem as the task of semantic segmentation and trained a fully convolutional neural network with skip connections to assign each pixel in the input image to one of the 30 predefined categories. Click in the video on the right to see some examples.

Impact

The implemented solution provides useful information which can help to identify drivable area and detect various traffic participants in the visual field of the board camera. The neural network may serve by providing additional confidence (due to redundancy) for a resulting perception system of a self-driving vehicle.