6533b7defe1ef96bd1275ab5

RESEARCH PRODUCT

Semantic Analysis of the Driving Environment in Urban Scenarios

subject

description

Understanding urban scenes require recognizing the semantic constituents of a scene and the complex interactions between them. In this work, we explore and provide effective representations for understanding urban scenes based on in situ perception, which can be helpful for planning and decision-making in various complex urban environments and under a variety of environmental conditions. We first present a taxonomy of deep learning methods in the area of semantic segmentation, the most studied topic in the literature for understanding urban driving scenes. The methods are categorized based on their architectural structure and further elaborated with a discussion of their advantages, possible limitations, and future directions. Then, we proposed a new approach to visual attention for driving based on a conditional generative adversarial network. We have presented the well-known salience algorithms, both classical and Deep Learning approaches, used for visual attention. We built a large visual attention database on a new strategy for mining saliency heatmaps from existing driving datasets. We then proposed a novel object identification framework that combines motion and geometry cues to understand the urban driving environment. A new moving object detection model is developed by integrating an encoder-decoder network with semantic segmentation and a disparity estimator. An image registration algorithm is proposed along with the optical flow to compensate for ego-motion. Extensive empirical evaluations on various driving datasets show that all the proposed methods achieve remarkable performance in terms of accuracy and demonstrate the effectiveness of the essential techniques for scene understanding in autonomous driving.