WS15: Multi-Sensor Fusion and Extended Object Tracking for Autonomous Vehicles

Workshop Code: 9i78c

Time: 13:30-17:00


  • Karl Granström
    Affiliation: Chalmers University of Technology, Sweden

  • Xinyu Zhang
    Affiliation: Tsinghua University, China

  • Pr Jean-Philippe LAUFFENBURGER
    Affiliation: Université de Haute-Alsace (Mulhouse)

Scope and Goals

Autonomous agents of different kinds are rapidly becoming more and more common, exemplified by the many autonomous car projects undertaken globally. A reliable environment perception is essential for modules like situation analysis and prediction, manoeuvre planning and trajectory planning. Typically, a setup of multiple sensors with complementary measurement principles and fields of view is used for environment perception. In order to handle missed detections and false alarms of the sensors, temporal filtering of the data using statistical multi-sensor fusion algorithms, e.g., multi-object tracking algorithms, or dynamic occupancy grid maps, is required.

The aim of this tutorial is to provide an overview of recent approaches for environment perception based on occupancy grid maps and multi-object tracking. Further, the audience is introduced to extended object tracking, with a focus on both the underlying theory and relevant real-world applications. The single-object tracking problem will be introduced briefly. Afterwards, two approaches for multi-sensor fusion are presented in detail: dynamic occupancy grid maps and multi-object tracking algorithms.

The multi-object tracking problem is typically defined as keeping track of an unknown number of moving objects, and historically it has been focused on so called point targets which give at most one detection per time step. However, modern radar or lidar sensors have increasingly higher resolution, meaning that it is common to see multiple detections per object. In order to be able to use point target algorithms for these sensors, heuristic clustering algorithms are applied to the raw measurements to obtain object hypotheses. In challenging scenarios, the hard decisions of the clustering algorithms affect the performance of the tracking algorithm due to the associated loss of information. Consequently, so called extended target tracking algorithms which are capable of handling several measurements per target are required.

In the tutorial part of this workshop, the modelling of object shapes and measurements in extended object tracking algorithms are introduced in detail and the integration in multi-object tracking algorithms is outlined. Finally, several applications of the presented extended object algorithms in automotive applications are shown.

Topics of Interest

  • Bayesian filtering
  • Multi-Object Tracking and Mapping
  • Multi-Sensor Fusion
  • Applications with data from cameras, lidars and radars
  • Tracking examples with pedestrians, vehicles, and cyclists


  • Bayesian Framework for Autonomous Vehicle Localization

  • Multi-Objective Adaptive Cruise Control Strategy Based on Variable Time Headway

  • Lane Detection and Road Surface Reconstruction Based on Multiple Vanishing Points

  • Low Latency V2X Applications and Network Requirements: Performance Evaluation