Vision based scene understanding for collision avoidance on roadway
Date of Issue2016
School of Computer Science and Engineering
Collision Avoidance Systems (CASs) are attracting a lot of attention as one of the most preferred solutions for advanced driver assistance and autonomous driving. However, scene understanding, which is an essential functionality in CASs, remains a major challenge mainly due to the need for real-time understanding of highly dynamic and complex environment. In this research, a number of robust and low complexity vision based scene understanding techniques for collision avoidance on roadway have been proposed. It has been well recognized in the literature that road surface detection in a dynamic environment is both challenging and computationally intensive. An efficient non-parametric road surface detection algorithm that exploits the depth cue is proposed to overcome the limitations of existing road surface detection methods. Unlike existing methods that attempt to fit the road surface into rigid models, the proposed method results in low computational complexity, mainly due to the reliance on four intrinsic road scene attributes observed under stereo geometry. It has been demonstrated that the proposed method is capable of detecting both planar and non-planar road surfaces. Extensive experimental results using three challenging benchmarks (i.e. enpeda, KITTI stereo/flow, and Daimler) show that the proposed road surface detection algorithm outperforms the baseline algorithms both in terms of detection accuracy (up to 23.12%) and runtime performance (up to 95.00%). Next, robust and low complexity algorithm for computing the ego-vehicle’s motion state is proposed. The proposed method estimates the ego-motion of the vehicle by first employing a novel pruning technique to reduce the computational complexity of the corner feature detection process without compromising on the quality of the extracted corner features. A robust and compute-efficient KLT tracker is proposed to facilitate the generation of the feature correspondences. Finally, an early RANSAC termination condition is introduced to the Gaussian-Newton optimization scheme to achieve rapid convergence of the motion estimation process. Evaluations based on the KITTI odometry benchmark show that the proposed visual odometry method outperforms the baseline algorithms both in terms of accuracy (up to 48.36%) and runtime performance. In addition, the proposed algorithm is placed among the top 15% when evaluated using the well-known KITTI odometry platform. Methods for robust and low complexity stereo-vision based obstacle detection and tracking are proposed. Unlike the works that focus only on the detection of vehicles or pedestrians, the proposed obstacle detection method relies on u-v disparity space to detect all obstacles in the scene. A Space of Interest (SOI) is defined to greatly reduce the search space of obstacles prior to employing adaptive hysteresis thresholding and connected component labeling techniques to segment SOI into sets of obstacles. Method for tracking obstacles across frames is also proposed by constructing a distinctive object appearance model. A number of strategies to further increase the distinctiveness and reduce the computational complexity for constructing the object model are also adopted. Finally, an online multi-object tracking framework is proposed by integrating the obstacle detection and data association modules in a robust way. Evaluations using the KITTI tracking benchmark confirm that the proposed obstacle detection and tracking method outperforms the baseline algorithm in terms of tracking accuracy by up to 51.78%. In addition, compared to the baseline algorithm that achieves about 0.23 frame per second (fps), the proposed method lends well for real-time performance with 20 fps. Finally, an efficient and robust risk assessment framework is proposed by integrating the obstacle detection and tracking, and visual odometry methods proposed in this thesis. The Extended Kalman Filter is customized to enhance the robustness of the predicted trajectory of the obstacles for assessing the collision risk. The robustness of collision prediction has been enhanced by accommodating positioning uncertainty. Evaluations based on the KITTI tracking dataset demonstrate that the proposed method are capable of robust and efficient assessment of the collision risk in diverse traffic scenarios. The proposed vision based scene understanding techniques in this research have paved the way towards realizing a real-time capable collision avoidance system that is both affordable and dependable.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision