Understanding Cityscapes: Efficient Urban Semantic Scene Understanding

Semantic scene understanding plays a prominent role in the environment perception of autonomous vehicles. The car needs to be aware of the semantics of its surroundings. In particular it needs to sense other vehicles, bicycles, or pedestrians in order to predict their behavior. Knowledge of the drivable space is required for safe navigation and landmarks, such as poles, or static infrastructure such as buildings, form the basis for precise localization. In this work, we focus on visual scene understanding since cameras offer great potential for perceiving semantics while being comparably cheap; we also focus on urban scenarios as fully autonomous vehicles are expected to appear first in inner-city traffic. However, this task also comes with significant challenges. While images are rich in information, the semantics are not readily available and need to be extracted by means of computer vision, typically via machine learning methods. Furthermore, modern cameras have high resolution sensors as needed for high sensing ranges. As a consequence, large amounts of data need to be processed, while the processing simultaneously requires real-time speeds with low latency. In addition, the resulting semantic environment representation needs to be compressed to allow for fast transmission and down-stream processing. Additional challenges for the perception system arise from the scene type as urban scenes are typically highly cluttered, containing many objects at various scales that are often significantly occluded. In this dissertation, we address efficient urban semantic scene understanding for autonomous driving under three major perspectives. First, we start with an analysis of the potential of exploiting multiple input modalities, such as depth, motion, or object detectors, for semantic labeling as these cues are typically available in autonomous vehicles. Our goal is to integrate such data holistically throughout all processing stages and we show that our system outperforms comparable baseline methods, which confirms the value of multiple input modalities. Second, we aim to leverage modern deep learning methods requiring large amounts of supervised training data for street scene understanding. Therefore, we introduce Cityscapes, the first large-scale dataset and benchmark for urban scene understanding in terms of pixel- and instance-level semantic labeling. Based on this work, we compare various deep learning methods in terms of their performance on inner-city scenarios facing the challenges introduced above. Leveraging these insights, we combine suitable methods to obtain a real-time capable neural network for pixel-level semantic labeling with high classification accuracy. Third, we combine our previous results and aim for an integration of depth data from stereo vision and semantic information from deep learning methods by means of the Stixel World (Pfeiffer and Franke, 2011). To this end, we reformulate the Stixel World as a graphical model that provides a clear formalism, based on which we extend the formulation to multiple input modalities. We obtain a compact representation of the environment at real-time speeds that carries semantic as well as 3D information