In a vehicle, data may come from very different sensors: camera, radar, IMU, GNSS, temperature sensor, oil pressure sensor, speed sensor, ... etc. In Gallopwave, we develop the technology to combine data from disparate sources to correct for the deficiencies of the individual sensors. Using the unified data, applications such as precise positioning, engine control, fault detection can be analyzed and fulfilled.
Having data from all different kinds of sources, deep learning is then applied to detect and recognize the patterns needed for the application. For example, lane and sign recognition for localization, vehicle detection and their path prediction for autonomous driving. However, the most challenging part is that the computation resources in the vehicle are limited. In Gallopwave, we reduce the computational complexity by model fusion and model redesign.
HD Map Construction
The traffic sign and lane, after being detected by the deep learning models, are reconstructed to 3D space using triangulation for HD map purpose. Moreover, big data processing is needed since the inputs are crowd-sourced and from different days and vehicles. We measure the quality of each path and aggregate them into a single map by optimizing the robust cost function.