A Shared Encoder DNN for Integrated Recognition and Segmentation of Traffic Scenes

Detection of traffic related objects in the vehicles surroundings is an important task for future automated cars. Visual object recognition and scene labeling from onboard cameras provides valuable information for the driving task. In computer vision, the task of generating meaningful image regions representing specific object categories such as cars or road area, is denoted as semantic segmentation. In contrast, scene recognition computes a global label that reflects the overall category of the scene. This contribution presents an efficient deep neural network (DNN) capable of solving both problems. The network topology avoids redundant computations, by employing a shared feature encoder stage combined with designated decoders for the two specific tasks. Additionally, element-wise weights in a novel Hadamard layer efficiently exploit spatial priors for the segmentation task. Traffic scene segmentation is examined in conjunction with road topology recognition based on the cityscapes dataset [2] augmented with manually labeled road topology data.