Abstract

In this paper, we present a deep neural network based real-time integrated framework to detect objects, lane markings, and drivable space using a monocular camera for advanced driver assistance systems. The object detection framework detects and tracks objects on the road such as cars, trucks, pedestrians, bicycles, motorcycles, and traffic signs. The lane detection framework identifies the different lane markings on the road and also distinguishes between the ego lane and adjacent lane boundaries. The free space detection framework estimates the drivable space in front of the vehicle. In our integrated framework, we propose a pipeline combining the three deep neural networks into a single framework, for object detection, lane detection, and free space detection simultaneously. The integrated framework is implemented in C++ and runs real-time on the Nvidia's Drive PX 2 platform.

Abstract

Nowadays, intelligent systems applied to vehicles have grown very rapidly; their goal is not only the improvement of safety, but also making autonomous driving possible. Many of these intelligent systems are based on making use of computer vision in order to know the environment and act accordingly. It is of great importance to be able to estimate the pose of the vision system because the measurement matching between the perception system (pixels) and the vehicle environment (meters) depends on the relative position between the perception system and the environment. A new method of camera pose estimation for stereo systems is presented in this paper, whose main contribution regarding the state of the art on the subject is the estimation of the pitch angle without being affected by the roll angle. The validation of the self-calibration method is accomplished by comparing it with relevant methods of camera pose estimation, where a synthetic sequence is used in order to measure the continuous error with a ground truth. This validation is enriched by the experimental results of the method in real traffic environments. This work was supported by the Spanish Government through the CICYTprojects (TRA2013-48314-C3-1-R and TRA2015-63708-R) and Comunidad de Madrid through SEGVAUTO-TRIES (S2013/MIT-2713). Publicado Document type: Article

Abstract

Wireless vehicular networks are to be deployed in both Europe and the USA within upcoming years. Such networks introduce a new promising source of information about vehicular environments to be used by cooperative advanced driver assistance systems (ADAS). However, development and evaluation of such cooperative ADAS is still challenging. Hence, we introduce a novel methodology for their development and evaluation processes. It is applied to evaluate the fulfillment of requirements on position accuracy information within exchanged messages. Such requirements are only roughly defined and not sufficiently evaluated in field tests. This holds especially for Global Navigation Satellite Systems (GNSS) optimized for maximum integrity of obtained positions. Such configuration is required to increase robustness and reliability of safety critical ADAS. We find that pure GNSS-based positioning cannot fulfill position accuracy requirements of studied ADAS in most test cases.

Abstract

The progress of safety technologies, based on the continuous advances in vehicle crash worthiness, restraint systems and active safety functions made traffic safer than ever before. Latest developments heading from assisted Advanced Driver Assistance System (ADAS) to Automated Driving (AD), lead to more and more complex real-world situations to be handled, going from standard driving tasks up to critical situations, which may cause a collision. Therefore, throughout the development process of such systems, it becomes common to use simulation technologies in order to assess these systems in advance. To gain results out of the simulation, input data are required; typically, from various sources, so the requirements can be covered. Thus, the challenge of scoping with different input sources arises. To come along with that problem, two main kinds of input data will be needed in general: (1) the descriptive parameter covering all border conditions, so called parameter room; (2) the system specifications for estimation. The quality of the results correlates strongly with the quality of inputs given. In case of ADAS systems and AD functions, the second kind of input data is very well known. Major challenges relate to the first kind of input data. Thus, the paper will describe a way to create input data that cover all descriptive parameters needed from normal driving up to the collision by the combination of accident analysis and real-world road traffic observations. The method aims at being applicable to different data sources and to different countries.

Abstract

Nowadays, driver assistance systems involve an ever increasing degree of automatization. Costly effort is put into testing the individual components to ensure proper functioning by the vehicle manufacturers. However, problems can also arise on a macroscopic scale, as vehicles and infrastructure are recently equipped with short range radio communication (“Car2X”). These problems caused by interaction are of even greater concern than “normal” bugs, as the final product might already have been deployed when the issues first become apparent. Multi-Agent System (MAS) research refers to such issues as “emergent misbehavior”. The said field also brought up an approach to automatically discover the worst consequences of the malfunctions. Hence, a given system under test can already be revised during development, saving a tremendous amount of resources. The approach from MAS is adapted to the domain of testing driver assistance systems in traffic simulations. A green-light optimal-speed advisory (GLOSA) algorithm is used as an example in which conceptual problems are discovered by the testing system after simpler issues are eliminated.

Abstract

Wireless vehicular networks are in the wake of mass deployment both in Europe and the USA. These networks introduce a new promising source of information about vehicular environments usable by cooperative advanced driver assistance systems (ADAS). However, development and evaluation of such ADAS is still challenging. Thus, we propose a methodology for their development and evaluation process. It is applied to evaluate the fulfillment of requirements on position accuracy information within the communicated data sets. Accuracy requirements are only roughly defined and not sufficiently evaluated in real world environments. This holds especially for GNSS (Global Navigation Satellite Systems) optimized for maximum integrity of obtained positions, which is required for safety critical ADAS to increase robustness and reliability. Our main goal is to determine whether position accuracy provided by GNSS is sufficient for cooperative ADAS. Thereby, we find that pure GNSS input cannot fulfill position accuracy requirements in most test cases.

Abstract

At an intersection with complex traffic flow, the early detection of the intention of drivers in surrounding vehicles can enable advanced driver assistance systems (ADAS) to warn the driver in advance or prompt its subsystems to assess the risk and intervene early. Although different drivers show various driving characteristics, the kinematic parameters of human-driven vehicles can be used as a predictor for predicting the driver’s intention within a short time. In this paper, we propose a new hybrid approach for vehicle behavior recognition at intersections based on time series prediction and deep learning networks. First, the lateral position, longitudinal position, speed, and acceleration of the vehicle are predicted using the online autoregressive integrated moving average (ARIMA) algorithm. Next, a variant of the long short-term memory network, called the bidirectional long short-term memory (Bi-LSTM) network, is used to detect the vehicle’s turning behavior using the predicted parameters, as well as the derived parameters, i.e., the lateral velocity, lateral acceleration, and heading angle. The validity of the proposed method is verified at real intersections using the public driving data of the next generation simulation (NGSIM) project. The results of the turning behavior detection show that the proposed hybrid approach exhibits significant improvement over a conventional algorithm; the average recognition rates are 94.2% and 93.5% at 2 s and 1 s, respectively, before initiating the turning maneuver. Document type: Article

Abstract

Driver inattention is one of the leading causes of traffic crashes worldwide. Providing the driver with an early warning prior to a potential collision can significantly reduce the fatalities and level of injuries associated with vehicle collisions. In order to monitor the vehicle surroundings and predict collisions, on-board sensors such as radar, lidar, and cameras are often used. However, the driving environment perception based on these sensors can be adversely affected by a number of factors such as weather and solar irradiance. In addition, potential dangers cannot be detected if the target is located outside the limited field-of-view of the sensors, or if the line of sight to the target is occluded. In this paper, we propose an approach for designing a vehicle collision warning system based on fusion of multisensors and wireless vehicular communications. A high-level fusion of radar, lidar, camera, and wireless vehicular communication data was performed to predict the trajectories of remote targets and generate an appropriate warning to the driver prior to a possible collision. We implemented and evaluated the proposed vehicle collision system in virtual driving environments, which consisted of a vehicle&ndash vehicle collision scenario and a vehicle&ndash pedestrian collision scenario. Document type: Article