Chih-Ming Hsu

Detecting drivable space in traffic scene understanding

Traffic scene understanding and perception is an important issue for intelligent vehicles and autonomous mobile robots. Especially in dynamic environments, the determination of drivable space and moving obstacles are fundamental requirement for road scene understanding. In this paper, we propose a vision-based approach combining road geometry and color features to percept road and moving obstacles in a dynamic environment from the camera mounted on the host vehicle. In the approach, a free road surface is detected first based on feature similarity search using statistical feature analysis (SFA) combined with a breadth-first search (BFS) algorithm to segment different intensity similarity regions in a road image. Then, the similarity between the road model (its color distribution) and the road region candidates is expressed by a metric derived from the Bhattacharyya distance. With the free road surface, the relative distance of preceding obstacles can easily be estimated using the obstacle scanning mechanism (OSM) and online camera calibration scheme. The experimental results have shown that the proposed approach can detect the drivable region and estimate the relative distance of preceding obstacles in real traffic scenes.

A Case Study on Highway Flow Model Using 2-D Gaussian Mixture Modeling

Traffic flow prediction is very important for real time information of travelers and dynamic route guidance system. In the past, various methodologies have been developed for traffic flow prediction. However, most of existing methods for the formulation of traffic flow are complex. In this paper, using a 2-D histogram projection method, a temporal-spatial traffic flow data is first reduced to a 2-D scatter plot and a 2-D Gaussian Mixture Modeling (GMM) is then used to estimate the traffic flow. This study shows that the traffic flow data can be simply represented by a linear combination of multiple Gaussian functions which demonstrates a good visualization of the temporal-spatial traffic flow data.

Optimal Multi-Sensor Selection for Driver Assistance Systems under Dynamical Driving Environment

This paper discusses the design and development of a multi-sensor selection algorithm for driver assistance systems. The multi-sensor selection problem is formulated as an optimization problem of integer linear programming. Both static and dynamic scenarios are considered in the formulation; where a static one denotes an off-line multi-sensor selection problem and a dynamic one denotes an on-line driving condition transformation problem. The objective for determining an optimal selection of sensors are in terms of guaranteed coverage, energy, bandwidth and reliability of sensors, etc. Based on a multi-sensor network architecture, different driving conditions, and driver commands, an optimal multi-sensor selection algorithm is developed for identifying the environment around the vehicle and warning the driver in real time. The selection algorithm and multiple-vehicle driving is implemented in a driving simulator software.

A Systematic Spatiotemporal Modeling Framework for Characterizing Traffic Dynamics Using Hierarchical Gaussian Mixture Modeling and Entropy Analysis

To accurately characterize traffic flow, a hierarchical Gaussian mixture modeling (GMM) framework is proposed for constructing a proper empirical dynamics model. The traffic flow data are first represented by a linear combination of multiple Gaussian functions for characterizing related timing and geographical parameters and for reducing the quantity of collected traffic data. To further examine dynamically changing behaviors, the phase-transition approach is used for identifying various traffic flow patterns and their dynamic switching behaviors. Furthermore, the information entropy on the traffic data collected at various vehicle detectors can be calculated for characterizing the location significance of these detectors. Detailed experimental analyses showed that five types of traffic flow patterns can be identified based on a six-month traffic data set from Taiwanese highway systems. Each traffic flow pattern indicates a distinct interpretation of a special dynamic traffic behavior.

Monocular vision-based drivable region labeling using adaptive region growing

The ability of intelligent vehicles to determine drivable region and perceive obstacles in dynamic environments is essential for maintaining safety and preventing accidents. In this paper, a vision-based drivable region labeling method is proposed. The method is based on an adaptive growing-based approach combining color features restrictions from an indicated drivable region in an efficient, stable, and precise method that can work in various scenes. The proposed method demonstrates that it distinguishes robustly and precisely between drivable region and non-drivable region in freeway, urban, rural road scenes with illuminant variance conditions, using color features restrictions estimated from indicated drivable region without specific machine learning algorithms.

Multi-sensor selection optimization and driver warning decision for dynamical virtual driving simulator

Abstract This article presents a framework for the design and development of a multi-sensor selection optimization mechanism for a driver assistance simulation system. The multi-sensor selection mechanism is formulated as an optimization problem and solved by integer linear programming. Both static and dynamic driving scenarios are considered in the formulation; where a static scenario denotes an off-line multi-sensor selection problem and a dynamic scenario denotes an online driving condition transformation problem. The objective function for determining an optimal set of sensors is in terms of guaranteed coverage, energy, bandwidth, and reliability of sensors. For different driving conditions, and driver commands, a selection algorithm is developed for identifying the environment around the vehicle and warning the driver. The selection algorithm and multiple-vehicle driving examples are successfully implemented in a virtual driving simulator.

Monocular vision-based range estimation of on-road vehicles

This paper presents a monocular vision-based range estimation of on-road vehicles approach. The proposed approach mainly combines non-drivable region from drivable region detection for detection region estimation instead of detecting the whole image, shadow detection for on-road object extraction, vehicle structure points estimation and adjusting for on-road vehicle classification, and motion vector and Kalman filter of on-road vehicles for collision avoiding. Extensive experimentation was performed to demonstrate that the proposed approach can correctly and dynamically estimate the relative distance of on-road vehicles in actual traffic conditions.

Characterizing highway traffic dynamics using GMM and phase transition analysis

Characteristic analysis of dynamical traffic flow is an important issue in traffic management and control systems. Especially for a dense-ramps highway system within urban area, there exists a highly nonlinear and sophisticated property of traffic flow dynamics. This paper proposes a Gaussian Mixture Modeling (GMM) to estimate dynamical traffic flow. The traffic flow data can be represented by a linear combination of multiple Gaussian functions to reduce the large amount of traffic data. Then the phase transition approach is adopted for analyzing various traffic flow characteristics. With the phase transition analysis, the dynamical switching behaviors of traffic flow can be characterized. The experimental results shown that the five types of traffic flow patterns are identified based on a real traffic dataset from Taiwan highway systems. Each traffic flow pattern can indicate a different interpretation of a special dynamical traffic behavior.

Traffic Flow Characteristic Based on A Phase Plane Approach

This paper proposes a phase-plane approach for analyzing the interpretation of various traffic flow characteristics. Using the phase-plane analysis, the dynamical switching behaviors of traffic flow can be easily studied. By analyzing the mathematical models of traffic flow transition, the dynamical behaviors can be characterized within the phase plane. Furthermore, 2-D Gaussian Mixture Modeling (GMM) is used to classify the dynamic traffic flow, and five types of traffic flow patterns are identified based on a real traffic dataset from Taiwan highway systems. Each traffic flow pattern can indicate a different interpretation of a special dynamical traffic behavior. A critical discussion of switching conditions for five types of flow patterns is also presented.