Enhancing Object Detection Capabilities of the Mapless Autonomy Platform with version 2023.03
driveblocks, a software provider for commercial vehicle autonomy, enhances the object detection and perception capabilities of its Mapless Autonomy Platform with the new version 2023.03. The platform includes algorithms for object detection, environment perception, sensor-fusion and trajectory planning in various commercial vehicle use-cases, such as highway driving, mining, container logistics or agricultural applications. It combines recent breakthroughs in artificial intelligence, such as transformer neural networks, with a geometrically interpretable approach for sensor-fusion and safety validation. The Autonomy Platform can be customized in a very data efficient way by refining the deep learning base models with customer and application specific data.
Major additions to object detection algorithms and data management workflow for machine learning components
The new release adds several new features and enhances the capabilities of already existing algorithms:
· A new LIDAR-based object detection and bounding box estimation algorithm
· Configurations for object detection in off-road use-cases, such as construction sites, mining and agriculture
· Integration of driveblocks deep learning components with a holistic data management pipeline deployed in the cloud, spanning the lifecycle from data collection up to model training & deployment.
Additional minor improvements of the platform include:
· Addition of a cabin motion compensation algorithm to achieve higher dynamic sensor-calibration accuracy
· Support for ROS1 bags in data management pipeline
· Significant compute time reductions in sensor-fusion to enable even more parallel perception pipeline data streams with similar compute hardware
· Generation of panorama images from LIDAR pointclouds providing different features such as intensity or distance as a basis for subsequent deep learning detection algorithms
The target of the Mapless Autonomy Platform is to speed up the path to production for vehicle OEMs, Tier-1s and Autonomous Driving companies by providing easy-to-use software components and reduced validation and certification efforts by re-using software components. Customers get access to a consistent environment model, including drivable space, potential driving corridors as well as stationary and dynamic object positions and sizes.