Search

For autonomous navigation, high definition maps are a widely used source of information. Pole-like features encoded in HD maps such as traffic signs, traffic lights or street lights can be used as landmarks for localization. For this purpose, they first need to be detected by the vehicle using its embedded sensors. While geometric models can be used to process 3D point clouds retrieved by lidar sensors, modern image-based approaches rely on deep neural network and therefore heavily depend on annotated training data. In this paper, a 2D HD map is used to automatically annotate pole-like features in images. In the absence of height information, the map features are represented as pole bases at the ground level. We show how an additional lidar sensor can be used to filter out occluded features and refine the ground projection. We also demonstrate how an object detector can be trained to detect a pole base. To evaluate our methodology, it is first validated with data manually annotated from semantic segmentation and then compared to our own automatically generated annotated data recorded in the city of Compi{\`e}gne, France. Erratum: In the original version [1], an error occurred in the accuracy evaluation of the different models studied and the evaluation method applied on the detection results was not clearly defined. In this revision, we offer a rectification to this segment, presenting updated results, especially in terms of Mean Absolute Errors (MAE).

Perception of other road users is a crucial task for intelligent vehicles. Perception systems can use on-board sensors only or be in cooperation with other vehicles or with roadside units. In any case, the performance of perception systems has to be evaluated against ground-truth data, which is a particularly tedious task and requires numerous manual operations. In this article, we propose a novel semi-automatic method for pseudo ground-truth estimation. The principle consists in carrying out experiments with several vehicles equipped with LiDAR sensors and with fixed perception systems located at the roadside in order to collaboratively build reference dynamic data. The method is based on grid mapping and in particular on the elaboration of a background map that holds relevant information that remains valid during a whole dataset sequence. Data from all agents is converted in time-stamped observations grids. A data fusion method that manages uncertainties combines the background map with observations to produce dynamic reference information at each instant. Several datasets have been acquired with three experimental vehicles and a roadside unit. An evaluation of this method is finally provided in comparison to a handmade ground truth.

For an autonomous vehicle, situation understand-ing is a key capability towards safe and comfortable decision-making and navigation. Information is in general provided bymultiple sources. Prior information about the road topology andtraffic laws can be given by a High Definition (HD) map whilethe perception system provides the description of the spaceand of road entities evolving in the vehicle surroundings. Incomplex situations such as those encountered in urban areas,the road user behaviors are governed by strong interactionswith the others, and with the road network. In such situations,reliable situation understanding is therefore mandatory to avoidinappropriate decisions. Nevertheless, situation understandingis a complex task that requires access to a consistent andnon-misleading representation of the vehicle surroundings. Thispaper proposes a formalism (an interaction lane grid) whichallows to represent, with different levels of abstraction, thenavigable and interacting spaces which must be considered forsafe navigation. A top-down approach is chosen to assess andcharacterize the relevant information of the situation. On a highlevel of abstraction, the identification of the areas of interestwhere the vehicle should pay attention is depicted. On a lowerlevel, it enables to characterize the spatial information in aunified representation and to infer additional information inoccluded areas by reasoning with dynamic objects.

We propose an information-fusion approach based on belief functions to combine convolutional neural networks. In this approach, several pre-trained DS-based CNN architectures extract features from input images and convert them into mass functions on different frames of discernment. A fusion module then aggregates these mass functions using Dempster's rule. An end-to-end learning procedure allows us to fine-tune the overall architecture using a learning set with soft labels, which further improves the classification performance. The effectiveness of this approach is demonstrated experimentally using three benchmark databases.

We propose a hybrid architecture composed of a fully convolutional network (FCN) and a Dempster-Shafer layer for image semantic segmentation. In the so-called evidential FCN (E-FCN), an encoder-decoder architecture first extracts pixel-wise feature maps from an input image. A Dempster-Shafer layer then computes mass functions at each pixel location based on distances to prototypes. Finally, a utility layer performs semantic segmentation from mass functions and allows for imprecise classification of ambiguous pixels and outliers. We propose an end-to-end learning strategy for jointly updating the network parameters, which can make use of soft (imprecise) labels. Experiments using three databases (Pascal VOC 2011, MIT-scene Parsing and SIFT Flow) show that the proposed combination improves the accuracy and calibration of semantic segmentation by assigning confusing pixels to multi-class sets., Comment: 34 pages, 21 figures

We propose a new classifier based on Dempster-Shafer (DS) theory and a convolutional neural network (CNN) architecture for set-valued classification. In this classifier, called the evidential deep-learning classifier, convolutional and pooling layers first extract high-dimensional features from input data. The features are then converted into mass functions and aggregated by Dempster's rule in a DS layer. Finally, an expected utility layer performs set-valued classification based on mass functions. We propose an end-to-end learning strategy for jointly updating the network parameters. Additionally, an approach for selecting partial multi-class acts is proposed. Experiments on image recognition, signal processing, and semantic-relationship classification tasks demonstrate that the proposed combination of deep CNN, DS layer, and expected utility layer makes it possible to improve classification accuracy and to make cautious decisions by assigning confusing patterns to multi-class sets.

The Limb Profiler (LP) is a part of the Ozone Mapping and Profiler Suite launched on board of the Suomi NPP satellite in October 2011. The LP measures solar radiation scattered from the atmospheric limb in ultraviolet and visible spectral ranges between the surface and 80 km. These measurements of scattered solar radiances allow for the retrieval of ozone profiles from cloud tops up to 55 km. The LP started operational observations in April 2012. In this study we evaluate more than 5.5 years of ozone profile measurements from the OMPS LP processed with the new NASA GSFC version 2.5 retrieval algorithm. We provide a brief description of the key changes that had been implemented in this new algorithm, including a pointing correction, new cloud height detection, explicit aerosol correction and a reduction of the number of wavelengths used in the retrievals. The OMPS LP ozone retrievals have been compared with independent satellite profile measurements obtained from the Aura Microwave Limb Sounder (MLS), Atmospheric Chemistry Experiment Fourier Transform Spectrometer (ACE-FTS) and Odin Optical Spectrograph and InfraRed Imaging System (OSIRIS). We document observed biases and seasonal differences and evaluate the stability of the version 2.5 ozone record over 5.5 years. Our analysis indicates that the mean differences between LP and correlative measurements are well within required +/- 10% between 18 and 42 km. In the upper stratosphere and lower mesosphere (>43 km) LP tends to have a negative bias. We find larger biases in the lower stratosphere and upper troposphere, but LP ozone retrievals have significantly improved in version 2.5 compared to version 2 due to the implemented aerosol correction. In the northern high latitudes we observe larger biases between 20 and 32 km due to the remaining thermal sensitivity issue. Our analysis shows that LP ozone retrievals agree well with the correlative satellite observations in characterizing vertical, spatial and temporal ozone distribution associated with natural processes, like the seasonal cycle and quasi-biennial oscillations. We found a small positive drift approx. 0.5%/yr in the LP ozone record against MLS and OSIRIS that is more pronounced at altitudes above 35 km. This pattern in the relative drift is consistent with a possible 100m drift in the LP sensor pointing detected by one of our altitude-resolving methods.

The theoretical basis of the Ozone Mapping and Profiler Suite (OMPS) Limb Profiler (LP) Version 1 aerosol extinction retrieval algorithm is presented. The algorithm uses an assumed bimodal lognormal aerosol size distribution to retrieve aerosol extinction profiles at 675 nm from OMPS LP radiance measurements. A first-guess aerosol extinction profile is updated by iteration using the Chahine nonlinear relaxation method, based on comparisons between the measured radiance profile at 675 nm and the radiance profile calculated by the Gauss Seidel limb-scattering (GSLS) radiative transfer model for a spherical-shell atmosphere. This algorithm is discussed in the context of previous limb-scattering aerosol extinction retrieval algorithms, and the most significant error sources are enumerated. The retrieval algorithm is limited primarily by uncertainty about the aerosol phase function. Horizontal variations in aerosol extinction, which violate the spherical-shell atmosphere assumed in the version 1 algorithm, may also limit the quality of the retrieved aerosol extinction profiles significantly.

One of the largest constraints to the retrieval of accurate ozone profiles from UV backscatter limb sounding sensors is altitude registration. Two methods, the Rayleigh scattering attitude sensing (RSAS) and absolute radiance residual method (ARRM), are able to determine altitude registration to the accuracy necessary for long-term ozone monitoring. The methods compare model calculations of radiances to measured radiances and are independent of onboard tracking devices. RSAS determines absolute altitude errors, but, because the method is susceptible to aerosol interference, it is limited to latitudes and time periods with minimal aerosol contamination. ARRM, a new technique introduced in this paper, can be applied across all seasons and altitudes. However, it is only appropriate for relative altitude error estimates. The application of RSAS to Limb Profiler (LP) measurements from the Ozone Mapping and Profiler Suite (OMPS) on board the Suomi NPP (SNPP) satellite indicates tangent height (TH) errors greater than 1 km with an absolute accuracy of +/-200 m. Results using ARRM indicate a approx. 300 to 400m intra-orbital TH change varying seasonally +/-100 m, likely due to either errors in the spacecraft pointing or in the geopotential height (GPH) data that we use in our analysis. ARRM shows a change of approx. 200m over 5 years with a relative accuracy (a long-term accuracy) of 100m outside the polar regions.

Following the successful launch of the Ozone Mapping and Profiler Suite (OMPS) aboard the Suomi National Polar-orbiting Partnership (SNPP) spacecraft, the NASA OMPS Limb team began an evaluation of instrument and data product performance. The focus of this paper is the instrument performance in relation to the original design criteria. Performance that is closer to expectations increases the likelihood that limb scatter measurements by SNPP OMPS and successor instruments can form the basis for accurate long-term monitoring of ozone vertical profiles. The team finds that the Limb instrument operates mostly as designed and basic performance meets or exceeds the original design criteria. Internally scattered stray light and sensor pointing knowledge are two design challenges with the potential to seriously degrade performance. A thorough prelaunch characterization of stray light supports software corrections that are accurate to within 1% in radiances up to 60 km for the wavelengths used in deriving ozone. Residual stray light errors at 1000nm, which is useful in retrievals of stratospheric aerosols, currently exceed 10%. Height registration errors in the range of 1 km to 2 km have been observed that cannot be fully explained by known error sources. An unexpected thermal sensitivity of the sensor also causes wavelengths and pointing to shift each orbit in the northern hemisphere. Spectral shifts of as much as 0.5nm in the ultraviolet and 5 nm in the visible, and up to 0.3 km shifts in registered height, must be corrected in ground processing.

Data assimilation is a potentially powerful method of extracting the maximum amount of information from observed data by combining their information with models. in this application, we will use stratospheric ozone profiles retrieved from EOS MLS and total column amounts retrieved from OMI' along with the GEOS-5 GCM, that includes a representation of ozone chemistry and transport. The presentation will discuss how well we can derive tropospheric ozone column information in this system, including its sensitivity to model errors and some assumptions made in the assimilation system. We discuss also ozone structures in the tropopause region and their sensitivity to model resolution and other fact=ors. An important issue we examine is how much more information we obtain from ozone assimilation than from model studies using state of the art chemistry models - answering such questions helps determine how useful ozone assimilation really is, given our present capabilities in chemical modeling.