Apr 21, 2021
Qualcomm products mentioned within this post are offered by Qualcomm Technologies, Inc. and/or its subsidiaries.
At Qualcomm Technologies, we’re focused on building autonomous driving (AD) technologies that can scale across all levels of autonomy and span from premium solutions to mass-market offerings. We believe that precise localization and robust perception are foundational to AD. Why? Autonomous vehicles mostly rely upon three technologies: high-precision localization, robust and reliable environment perception (including both dynamic and static objects), and the prediction and motion planning algorithms for safe maneuvers. The higher the accuracy of the localization and perception data, the easier and more reliable is the motion planner to achieve required autonomous driving goals.
Our unified localization solution
We take a holistic and modular approach when solving localization challenges for AD. An autonomous vehicle must be able to localize itself within a few centimeters of accuracy to plan and execute complex maneuvers required in real-world road conditions. HD maps are increasingly becoming a preferred input to localization systems to achieve such high levels of accuracy. So, any localization solution that’s deployed must be robust to HD map errors that can arise due to changing road geometries. Our approach is to develop advanced and optimized modules that solve such challenges for our automotive OEM customers and partners.
One such component is Qualcomm Vision Enhanced Precise Positioning (VEPP) software, which can provide lane-level positioning and a precise heading in the global coordinates by tightly fusing GNSS, IMU, camera, and vehicle odometry information in real time and in virtually all environments. Our localization solution fuses the global pose with camera perception and HD map information as well as provides an ability to generate a local map, producing identification of gaps or errors in HD maps on-the-fly — all while the vehicle can operate in an autonomous mode. This unique capability allows for vehicles equipped with our localization modules to contribute toward improving and maintaining the accuracy of HD maps over time.
Novel techniques for efficient environment perception
Scaling environment perception technologies efficiently has been one of our primary research focus areas. Our goal is to harness all sensor types including cameras, radar, and lidar, as well as maps to build novel ways of solving environment perception challenges.
One of the key areas within perception is multisensor data fusion, which includes data association and object-state estimation. In the context of AD, data association is the ability of the autonomous vehicle to reliably correlate data from different views and modalities of sensors. State estimation involves tracking of real-world objects (like vehicles, pedestrians, lanes, curbs, traffic signs, etc.) over time and estimate their size, speed, and position relative to the autonomous vehicle itself. We have achieved impressive results showing how we could solve complex data association problems using advanced camera and radar processing algorithms.
Another area of focus for us has been to accurately predict the pose and track the size of objects on the road with 3D estimations. We are using 2D data from cameras, radar, and maps to build these 3D estimations leveraging both deep learning and concepts from multi-view geometry. Such accurate 3D estimations can enable human-like maneuvers for advanced safety of the autonomous vehicle.
This is only a portion of what the team works on every day as we continue to push the limits of the state-of-the-art technologies that solve AD challenges.
Accelerating autonomous driving solutions
Qualcomm Technologies has been at the forefront of providing connectivity solutions to the point that our telematics solutions are widely accepted throughout the auto industry. Building on this solid understanding of the automotive industry, we are now helping drive some of the biggest innovations in AD. Our comprehensive frameworks and tools around cross-sensor calibration, active learning, big data management, and automated verification/validation can significantly reduce the time-to-market for autonomous vehicle manufacturers.