DRIVE Labs: How Localization Helps Vehicles Find Roads

2020-01-17

Localization is a key feature of autonomous vehicles, allowing them to accurately position themselves within centimeters on a map. This high level of accuracy enables self-driving cars to understand their surroundings and establish a sense of road and lane structure. With localization, vehicles can detect when lanes diverge or merge, plan lane changes, and determine lane paths. DRIVE Localization achieves precise positioning by matching semantic landmarks in the vehicle's environment with features in a high-definition map. In this DRIVE Labs, we demonstrate how our localization algorithm enables high precision and robustness using sensors. Localization provides 3D motion for autonomous vehicles within a high-definition (HD) map, including 3D position, 3D direction, and their uncertainties. Unlike GPS navigation maps, autonomous vehicles require higher precision in positioning relative to the map, typically within a few centimeters. While the value of localization to HD maps may not be immediately apparent due to the rapid development of real-time perception, it allows cars to make decisions based on information beyond their current field of view. For example, the map can signal to the car when the current lane will terminate in certain directions. The car can then exit the lane at any available space and merge into an adjacent lane, avoiding potential hazards. Localization to HD maps also enables self-driving cars to contextualize the behavior of other nearby vehicles in the map. For instance, at an intersection, it can identify oncoming cars in the left-turn lane and safely execute an unprotected left turn. Achieving precise localization typically involves expensive inertial sensors, GNSS sensors, and lidar. However, to popularize sensors in the mass market, such as cameras and low-cost inertial GNSS and CAN sensors, innovation is required. DRIVE Localization considers various potential feature correspondences to accommodate data from sensors with less accuracy and more noise from the mass market. By overlaying the detection results from camera data onto the HD map, candidate actions can be scored. The module evaluates thousands of candidate actions in parallel, scoring the match between map features and a frame of visual data. To address the ambiguity caused by noisy feature detection, the module also relies on temporal cross-checking of actions and employs filters. The computational requirements of this algorithm are demanding, but the design is tailored specifically for NVIDIA Xavier system-on-chip and GPU parallel processing, which enables efficient processing on the NVIDIA DRIVE AGX platform. The robustness of our algorithm allows it to work in challenging environments such as nighttime, rainy and foggy conditions, and GNSS-denied areas like tunnels, even with just a single camera. DRIVE Localization can be used independently with one or multiple cameras, radars, or lidars. Each sensor modality provides accurate outputs that can be directly used for localization. Subsequently, the localization from these individual outputs can be fused to provide a single, even more reliable output. The ability to use multiple sensors provides additional safety by checking the consistency between different localization outputs. For example, if one sensor fails calibration or is severely affected by adverse weather conditions like rain or snow, the localization results from these sensors will differ significantly from other sensors. DRIVE Localization can detect such anomalies and generate fusion results only from consistent sensor outputs. By leveraging this type of redundancy, self-driving cars can achieve higher levels of safety. In addition to camera localization, DRIVE Localization can also be used in conjunction with radar (top) and lidar (bottom). The top image shows real-time radar responses as blue dots, while the bottom image displays the reflectivity image of the lidar and real-time lidar points in green. An example of DRIVE Localization is the simultaneous fusion of localization results from cameras, radar, and lidar.

RECENT AI NEWS

RECENT AI TOOLS