A Review on Recent Encoder-Decoder Architectures for Drivable Region Detection and Segmentation

With the proliferation of science and technology, autonomous driving has garnered significant attention in the recent years due to the manifold benefits of driverless vehicles, including enhanced quality of life, mitigation of accidents, and curtailment of traffic congestion. In this work, recently developed encoder-decoder oriented architectures specifically designed to detect and segment the safe drivable region on roads have been reviewed, analyzed and compared, and the challenges in this field have been discussed. In pursuit of this objective, we initially furnish a survey of various orthodox and deep learning methods that find their use in detection of the road obstacles, and segmentation of roads and lane-lines. This is followed by a discussion on relevant field-oriented datasets. Subsequently, we present some recent encoder-decoder based methodologies such as HybridNets, You Only Look Once Panoptic (YOLOP), etc. These methods are then compared in terms of both inference speeds as well as performance metrics, to understand their potential in the real-time better, concluding that YOLOPv2 is currently the best suited architecture meant for drivable region detection. This is followed by an examination of challenges and unresolved inquiries in this domain.