HERE Technologies provides real-time location services that enable people, enterprises, and cities around the world to harness the power of location and create innovative solutions for a safer and more efficient living. Multimedia retrieval techniques and sensor fusion approaches are essential for enriching location services and for keeping the underlying map up to date. In this talk, I will give an overview of some of the work we do in the CTO Research group to support existing location services and enable future ones. We aim to automatically extract useful information from massive collections of images, LiDAR point clouds, car sensor data and open web data. I will present work related to image recognition for map making purposes, information retrieval for points of interest enrichment, and work related to creating a highly accurate map of the roads and cities for the future autonomous navigation services.

In this paper we present our experience on detecting and classifying traffic jams in real time from probe data. We classify traffic jams at two levels. At a higher level, we classify traffic jams into recurring and non-recurring jams. Then at a lower level we identify accidents out of non-recurring jams based on features that characterize upstream and downstream traffic patterns. Accidents are highly unpredictable and usually create heavy and long lasting congestion, and therefore are particularly worth detecting. We discuss the challenges of detecting accidents in real time as well as our approaches and results.

This paper proposes an algorithm called MACH for coding a hierarchical category structure to enable prefix matching between a publication and a subscription. Compared to fixed-length coding, the MACH algorithm generates shorter code lengths without losing the efficiency of matching. In addition, the MACH algorithm automatically identifies the bits that are not necessary for prefix matching. These bits are saved to allow growth of categories. The bit saving is prioritized according to a configurable criterion. Experiments with real world data and synthetic data are conducted to evaluate the benefit of the proposed algorithm.

Modern object detection methods apply binary linear classifiers on Euclidean feature vectors. This paper shows that projecting feature vectors onto a hypersphere allows an approximate range query to accelerate these detectors within acceptable levels of accuracy. The expense of constructing the k-d tree used by these range queries is justified when many detectors are used. We demonstrate our acceleration technique on several existing detection systems, including a state of the art logo detector, and show that approximate range queries can detect logos at least half as well at 11× the speed of the full fidelity method.

While there is significant recent progress in technologies supporting augmented reality for small indoor environments, there is still much work to be done for large outdoor environments. This workshop focuses primarily on research that enables high-quality outdoor Mixed Reality (MR) and Augmented Reality (AR) applications. These research topics include, but are not restricted to: — 3D geo-referenced data (images, point clouds, and models) — Algorithms for object recognition from large databases of geo-referenced data — Algorithms for object tracking in outdoor environment — Multi-cue fusion to achieve improved performance of object detection and tracking — Novel representation schemes to facilitate large-scale content distribution — 3D reasoning to support intelligent augmentation — Novel and improved mobile capabilities for data capture (device sensors), processing, and display — Applications, experiences, and user interface techniques. The workshop will also showcase existing prototypes of applications enabled by these technologies: mirror worlds, high-fidelity virtual environments, applications of panoramic imagery, and user studies relating to these media types. This workshop aims to bring together academic and industrial researchers and to foster discussion amongst participants on the current state of the art and future directions for technologies that enable large-scale outdoor MR and AR applications. The workshop will start with a session in which position statements and overviews of the state of the art are presented. In the afternoon, we will follow up with discussion sessions and a short closing session.

We introduce a new class of mobile augmented reality navigation applications that allow people to interact with transit maps in public transit stations and vehicles. Our system consists of a database of coded transit maps, a vision engine for recognizing and tracking planar objects, and a graphics engine to overlay relevant real-time navigation information, such as the user's current location and the time to destination. We demonstrate this system with a prototype application built from open source components only. The application runs on a Nokia N900 mobile phone equipped with Maemo, a Debian Linux-based operating system. We use the OpenCV library and the new Frankencamera API for the vision engine. The application is written using the LGPL licensed Qt C++ Framework.

This paper presents a novel method to process large scale, ground level Light Detection and Ranging (LIDAR) data to automatically detect geo-referenced navigation attributes (traffic signs and lane markings) corresponding to a collection travel path. A mobile data collection device is introduced. Both the intensity of the LIDAR light return and 3-D information of the point clouds are used to find retroreflective, painted objects. Panoramic and high definition images are registered with 3-D point clouds so that the content of the sign and color can subsequently be extracted.

In this paper, we present the image retrieval and computer vision applications implemented for massive mobile image data collection at NAVTEQ. This novel and comprehensive mobile data collection method captures ground level imageries around the world. Automatic image analysis and modeling provides NAVTEQ with next-generation map making capabilities. This allows NAVTEQ to exceed customers' expectations by creating and maintaining digital maps that are more timely, complete, accurate, and which contain richer content.

Clip art is a simplified illustration form consisting of layered filled polygons or closed curves used to convey 3D shape information in a 2D vector graphics format. This paper focuses on the problem of direct conversion of smooth surfaces, ranging from the free-form shapes of art and design to the mathematical structures of geometry and topology, into a clip art form suitable for illustration use in books, papers and presentations. We show how to represent silhouette, shadow, gleam and other surface feature curves as the intersection of implicit surfaces, and derive equations for their efficient interrogation via particle chains. We further describe how to sort, orient, identify and fill the closed regions that overlay to form clip art. We demonstrate the results with numerous renderings used to illustrate the paper itself.