4.5.3 Camera Calibration
We now present another application of the SMRF algorithm, in the context of camera calibra- tion. The goal is to accurately estimate the position and orientation of the camera with respect to the road and to its intrinsic parameters. A calibration pattern made of two sets of perpen- dicular lines painted on the road is observed by a camera mounted on a vehicle, as shown in Figure 4.5(a). The SMRF algorithm can be used to provide accurate data to the calibration algorithm by estimating the grid intersections. Even though the markings are clearly visible in the image, some of them are quite short, and there are outliers due to the presence of water puddles. Figure 4.5(b) shows the extracted lane-marking centers. When a Gaussian mixture model is used, the obtained fit is severely troubled by the outliers, as displayed in Figure 4.5(d), even though the curves are initialized very close to the expected solution, see Figure 4.5(c).
For autonomous navigation, the only information of the road space is, however, not sufficient because vehicles need to know where it is authorized to go, and so, lane-level information is required. When there is an obstacle in the ego- lane, the host vehicle has two options: i) keep-lane and stop or ii) lane-change. The second alternative needs semantic road rule information to evaluate if the space is accessible or not. According to road rules, solid lane markings are normally forbidden to cross, whereas dashed ones indicate possible lane change. These lanemarking types imply the road rules that should be obeyed by the host vehicle. Many lane detection methods have been studied and developed [4][5][6] for this purpose. Nevertheless, lane detection methods relying 1 The authors are with Sorbonne Universités, Université de Technologie
4.1 Quantitative Evaluation
The lanemarking detection algorithm described in this paper has been imple- mented on a computer that had an Intel i7, 2.40 GHz CPU. The algorithm was executed in Visual C++ with OpenCV. The processing time was approximately 25 ms. per frame in good and complex road condition. As part of the evaluation of our approach, we have images from the “ROMA” database. It comprises more than 100 Heterogeneous images of various road scenes. Each image is delivered with a manually constructed ground truth, which indicates the position of the visible road markings.
4.1 Quantitative Evaluation
The lanemarking detection algorithm described in this paper has been imple- mented on a computer that had an Intel i7, 2.40 GHz CPU. The algorithm was executed in Visual C++ with OpenCV. The processing time was approximately 25 ms. per frame in good and complex road condition. As part of the evaluation of our approach, we have images from the “ROMA” database. It comprises more than 100 Heterogeneous images of various road scenes. Each image is delivered with a manually constructed ground truth, which indicates the position of the visible road markings.
y = a
x − h + b(x − h) + v. (1)
In (2), x , y are the coordinates in pixels of the road marking points, a is the curvature parameter of the hyperbola, b is the slope of straight lanemarking, and (h,v) is the coordinate vector of the vanishing point. An ap- proach has been proposed in Tan et al. (2014), which uses a hyperbola model, with an improvement by proposing a ”River Flow” Lim et al. (2012) to detect curved lanes in difficult conditions, including dashed line markings and vehi- cle occlusion. In order to determine the curvature coefficient, a new method called ”Improved River Flow”Tan et al. (2015) was proposed to search the characteristic points in the far field that corresponds to the lanemarking. In Jang et al. (2014), another kind of approaches is proposed based on the geometrical model. The general idea of the method is to apply two main steps: the first one involves the extraction of the road characteristics, and the second step relies on a geometrical model for lane detection. the candidate lines are detected by applying a voting system to extract the one that better fits the lanes. In Chen and He (2012), the authors proposed a method to detect sharp curve lanes using the maximum likelihood principle.
III. L ANE MARKING EXTRACTION USING LOCAL THRESHOLD ALGORITHMS
In most algorithms, the extraction stage processes image lines independently, in a sequential fashion, seeking seg- ments that might correspond to lanemarking elements. Road markings are white (more rarely, colored, but most of the time light) strips of constant normalized width, painted on the dark roadway. Hence, extraction algorithms exploit, more or less fully, the characteristics of road markings with respect to photometry, geometry or both. To make use of the contrast between markings and pavement, a detection threshold, T G , is put on the intensity. From a geometric point of view, the constant width assumption is often violated due to wear, dirt and occlusions. Hence, instead of a threshold, an acceptability range is defined. Since markings are observed in perspective, this range must be adapted to the vertical position in the image (i.e. the line number, v). We will then denote it by [S m (v), S M (v)]. In the following evaluation on the Mitowns database, these parameters are fixed to same values for all the extractors. S m is linearly interpolated with value 1 at the line of horizon and with value 35 at the bottom of the image. To extract all the large markings, S M is linearly interpolated with value 1 at the line of horizon and with value 350 at the bottom of the image. These value were obtained by testing different values and by choosing the ones maximizing the DSC.
y = a
x − h + b(x − h) + v. (1)
In (2), x , y are the coordinates in pixels of the road marking points, a is the curvature parameter of the hyperbola, b is the slope of straight lanemarking, and (h,v) is the coordinate vector of the vanishing point. An ap- proach has been proposed in Tan et al. (2014), which uses a hyperbola model, with an improvement by proposing a ”River Flow” Lim et al. (2012) to detect curved lanes in difficult conditions, including dashed line markings and vehi- cle occlusion. In order to determine the curvature coefficient, a new method called ”Improved River Flow”Tan et al. (2015) was proposed to search the characteristic points in the far field that corresponds to the lanemarking. In Jang et al. (2014), another kind of approaches is proposed based on the geometrical model. The general idea of the method is to apply two main steps: the first one involves the extraction of the road characteristics, and the second step relies on a geometrical model for lane detection. the candidate lines are detected by applying a voting system to extract the one that better fits the lanes. In Chen and He (2012), the authors proposed a method to detect sharp curve lanes using the maximum likelihood principle.
Abstract— Localizing the vehicle in its lane is a critical task for any autonomous vehicle. By and large, this task is carried out primarily through the identification of ego-lane markings. In recent years, ego-lanemarking detection systems have been the subject of various research topics, using several inputs data such as camera or lidar sensors. Lately, the current trend is to use high accurate maps (HD maps) that provide accurate information about the road environment. However, these maps suffer from their availability and their price tag. An alternative is the use of affordable low-accurate maps. Yet, there is relatively little work on it. In this paper, we propose an information-driven approach that takes into account inaccurate prior geometry of the road from OpenStreetMap (OSM) to perform ego-lanemarking detection using solely a lidar. The two major novelties presented in this paper are the use of the OSM datasets as prior for the road geometry, which reduces the research area in the lidar space, and the information-driven approach, which guarantees that the outcome of the detection is coherent to the road geometry. The robustness of the proposed method is proven on real datasets and statistical metrics are used to highlight our method’s efficiency.
Other researches use the lane assignment information of different vehicles that communicate with each others [8]. The GPS information shared between these vehicle is used to calculate a probability for ego-lane determination. Al- ternative approaches use the on-board sensors for ego-lane estimation. The author in [9] estimates the probability of belonging to a lane, using lane change information and lane- markings detector. A similar approach is proposed in [10]. However, the results of lane-marking feed a Bayesian Net- work which is temporally filtered by a particles filter. An interesting approach is presented in [11], where the authors introduce a Bayesian Network for ego-vehicle localization in intersections. The Bayesian Network takes as an input the information from a sensorial perception system and a priori digital map. The approach shows interesting results. However, the dynamic of the vehicle is not considered, hence no temporary relation between frames is taken into account. In [12] Ego-lane localization is achieved from multiple- lanes detection. First, the author identifies the own-lane geometry, then adjacent lanes are hypothesized and tested, assuming same curvature and lane’s width. A more recent approach [13] fuses the position of surrounding vehicles with a map and a lane-marking detector into a Bayesian filter. The early tests show promising results when surrounding vehicles are detected. However, real-world experimental results are missing to assess the efficiency of the approach especially when there is no surrounding vehicle. Ego-lane estimation can be formulated as a scence classifcation problem, as in [14], where the authors describes the scene in a holistic way bypassing individual object detection: {vehicle, lane markers}.
High integrity lane level localization using multiple lane markings detection and horizontal protection levels
Gabriel Frisch, Philippe Xu and Emmanuel Stawiarski
Abstract— For autonomous driving, lane level accurate local- ization is a necessity for complex driving maneuvers. Classical GNSS based methods are usually not accurate enough to have an unambiguous lane level localization. Having camera measurements such as lanemarking detections along with high definition maps can enhance localization performance. In this paper, we are interested in high integrity localization, meaning being robust with low risk level. We propose a novel geometrical approach using horizontal protection levels on localization to propagate uncertainties and use lane markings to have an unambiguous map-matching. We demonstrate on real data that the algorithm can cope with high levels of noise on both localization and detection.
Chapter 3
Combining Vision with LiDAR
3.1 Motivation
Image data lacks the inherent physical properties of LiDAR data: a single pixel, unlike a single LiDAR point, says nothing about the car’s environment, absent other context. As a result, one can only perform limited checks on the lane lines using vision alone. In a particular case involving a Tesla driving with autopilot, the vehicle almost crashed into a concrete barrier. The problem was that bright lines that look similar to lane lines appeared on the barrier in the camera’s image [42]. These lines were actually bands of direct or reflected sunlight. In most cases, we believe that the vision tests described above would catch such anomalies; we took one particular image from an online video (Fig. 3-1) illustrating this problem and confirmed that the inferred lane lines would indeed fail the geometric test and be correctly rejected. However, it is conceivable that these spurious lane lines would have passed the geometry and conformance tests. Indeed, in the Tesla example, the yellow ray of sun almost looked, in the image, like a plausible lane line. Apparently a more basic property is being violated: the detected lane line is on the barrier and not on the ground.
Several improvements have been proposed to increase the capacity of Marking Menus while keeping their efficiency, accuracy and learnability. Zhao and Balakrishnan [10] im- proved the expert mode of hierarchical Marking Menus with a design where the user draws several marks instead of a sin- gle compound one. However, even if this design increases the potential depth of hierarchical menus, breadth remains a limiting factor for novice users and for menus with many items. Zone and Polygon Menus [9] address this issue with an alternate design that increases the breadth up to 16 items. However, Bailly et al. [1] pointed out that these improved designs have only been evaluated in expert mode, and raised the hypothesis that the transition from novice to expert may
The paper is organized as follows. Section 1 constitues the background for DOM analysis. In sections 2 and 4, we point out some aspects of the syntax-semantics of bare nouns (BNs), more precisely we examine some differences between bare singulars (BSs) and bare plurals (BPs) especially when they occur as direct objets. In sections 3 and 4, we examine the relationship between differential marking and BNs when the latter occur as direct objects. In section 6, we provide an analysis for the contrast mentioned above, while in section 7 we offer additional Spanish data illustrating the phenomenon. Finally, section 8 provides the conclusions of our research and announces a number of issues for further research.
VII. C ONCLUSION
MM integrity is becoming an important topic for intelligent vehicles that use information obtained from maps in safety applications. In this paper we have presented a PF-based MM algorithm developed to retain all the likely matching hy- potheses throughout the map-aided dead-reckoning, and which includes an integrity check procedure that uses redundant GNSS positions to detect faults. Using real road data, the filter was shown experimentally to respect the integrity of its solution set (i.e. always returning a set containing the correct matching). A new method for internal integrity monitoring was developed using a Mahalanobis distance as decision variable for detecting faults: the distance of the GNSS fix to each matching hypothesis is computed in order to test the coherence of the map-matching. The computation of the covariance matrices of the hypotheses obtained from the particle set is quite straightforward and well adapted to this computation. This allows the positioning system to be flagged as either “Use" or "Don’t Use” at a given time step. Experimentally, this system provides only 0.54% of misdetections and has a 65.6% availability (correct hypothesis tagged as “Use”) with the low cost sensors used in the experiments. While the algorithm may be over-pessimistic in difficult conditions (e.g. a dense urban environment with multi-lane roads), it is a good indicator for determining whether the positioning system requires complementary information to make it reliable (e.g. multipath detection and correction).
The major challenge in designing such a system is that it requires a procedure that allows to identify the system states from which there exists an admissible steering input that can keep the vehicle in the lane. The set of these states is called controlled invariant safe set. Determining whether a state is in this set (a safety verification task) often requires computationally intensive algorithmic procedures. Therefore the application of controlled invariance to real- world engineering problems has been limited to systems with special structures or small dimensions, see [3], [9], [10], [14] and references therein. Here we show that under certain conditions, that we specify, the controlled invariant safe set has a simple characterization that can be computed efficiently online. We therefore propose a system architecture that performs an initial safety check at the time the driver requests activation of the LDAS. If activation is safe (based on the above mentioned conditions), the system is enabled and continuously monitors the driver’s steering inputs. The driver’s inputs are overridden only if necessary to keep the vehicle in the lane.
• temporal verbs referring to the cycle of the sun (often subsumed under meteorological
predications, e.g. in Malchukov & Siewierska 2011) sometimes (NOT always) carry non-visual evidential marking
• however, the events referred to are clearly visible to the speaker why?
(e) Un buen sablazo de sol traspasaba (*a) la sacristia.
A quick examination of this construction shows that, in spite of its inanimate features (cf. parameter (i)), the direct object la sacristia ‘the vestry’ may be marked by a (as in (d)) due to its left dislocated position (and maybe to a special accentuation or to clitic doubling). Moreover, if the direct object stays in its in situ position, i.e. the post verbal position (as in (e)), it does not trigger marking altough we would expect it since the noun has a referential / specific reading (cf. parameter (ii)).
Aiming to describe realistic situations in detail, we allow for the speed laws and the number of lane to change along the road. In the study, for sake of simplicity, we consider the model proposed in [8], but more general source terms could be taken into account.
We consider an infinite road described by the real line. Let M ` ⊂ N + be the set of
In this paper I argue that despite their strong similarity, the focus construction in (2) is a simple clause, whose focus-marking effect results from the non-prototypical employment of the noun as predicate and of the verb as part of the argument phrase. The construction in (3), by contrast, is a cleft. Its matrix clause is composed of a pronoun and a nominal predicate, which together constitute a full-fledged equational clause: ‘It is/was the/a jaguar.’ 4 The argument phrase containing the verb corresponds to the relative (or “relative-like”, Lambrecht 2001: 467) clause that is included in most, if not all, definitions of clefts proposed in the literature. The two constructions also differ in their pragmatic function: The cleft is a specificational sentence, whereas the nominal-predicate focus construction is a simple predication, whose marked status comes from the non-prototypical pragmatic employment of noun and verb.
[7] Z. Dahmani, M.Z. Sarikaya: On a generalized Lane-Emden fractional dif- ferential system and its ∆−stability, Accepted on IASR-JARDCS of Dy- namical and Control Systems. Accepted End 2015.
[8] J. Dvila, L. Dupaigne, J. Wei: On the fractional Lane-Emden equation, Transactions of the American Mathematical Society, Vol. 369(9), 2014. [9] H. Fenga, C. Zhai: Existence and uniqueness of positive solutions for a