Semantic Scene Understanding in Unstructured Environment with Deep Convolutional Neural Network

Number of road fatalities have been continuously increasing since last few decades all over the world. Nowadays advanced driver assistance systems are being developed to help the driver in driving process and semantic scene understanding is an essential task for it. Convolutional Neural Networks (CNN) have shown impressive progress in various computer vision tasks including the semantic segmentation. Various architectures have been proposed in literature but loss of spatial acuity in semantic segmentation prevents them from achieving better results as details of small objects are lost in downsampling. To overcome this drawback, we propose to use dilated residual network as backbone in DeepLabV3+ which enables to preserve the details of smaller objects in the scene without reducing the receptive field. We focus our work on India Driving dataset (IDD) containing data from unstructured traffic scenario. Proposed architecture proves to be effective compared to earlier approaches in literature with 0.618 mIoU.