Executive summary
Advanced Driver Assistance Systems (ADAS) is revolutionary technology that could change the nature of transportation by increasing car and driving safety. The scope of ADAS is huge:
ADAS incorporates dozens of technologies, including:
- Lane Keeping Assist
- Braking Assist
- Blind Spot Indication
- Adaptive Cruise Control
- Parking Assist
About a third of all new cars sold in major markets have ADAS features, and that number is only growing
ADAS relies on the fact that the system can accurately perceive its environment and respond to it. In order to do its job, ADAS relies on data on the vehicle’s surroundings, and lots of it. This data is constantly streaming in through a handful of sensors on the vehicle, such as numerous cameras, radars, sonars, and perhaps most importantly, light detection and ranging (LiDARs).
This data needs to be of high quality for the algorithms to be accurate to ensure safety.
AVL Challenge
Perception of the environment is both a crucial and a challenging task for Advanced Driver Assistant Systems (ADAS) and Autonomous Driving (AD). A wide range of sensors, including LiDAR, cameras and radars are used to build the perception layer of these systems and ensure a precise view of the surroundings in diverse traffic and environmental situations.
In order to enhance data accuracy for ADAS & AD systems, AVL has developed Dynamic Ground Truth:
- High resolution measuring system used to capture reliable environment data
- Box mounted on vehicle w/ lidar, camera, and high precision GPS sensors
- Lets you capture 360 degree field of view.
To advance AVL’s Dynamic Ground Truth reference system, AVL needed two key things:
- Quick Sensor Calibration
- High accuracy labelling, delivered fast
AVL’s problem statement
Data integrity is at core of AVL’s DGT system. AVL’s problems were two fold:
Traditional sensor calibration has many inherent issues:
- Time consuming
- Expensive
- Not scalable
Existing Data annotation tools and services providers were not good enough:
- Inexperienced labeling team
- Lack of custom engineering
- Poor data operations and lifecycle management
Deepen AI ensures data integrity
Step 1 - Deepen Calibrate - High quality data collection through calibrated sensors
Deepen accomplishes this through its web-based calibration suite
- If sensors aren’t calibrated → then all the data they collect is less accurate → everything the data are used for (e.g. teaching the machine) becomes less reliable
- Deepen calibrates a variety of sensors (LiDAR, radar, camera, etc) to quickly provide a complex, detailed, and accurate reading of the data used.
Deepen’s calibration suite cuts the time spent on multi-sensor calibration from hours to seconds.
Step 2 - Deepen Annotation - Accurate data annotation
Deepen’s annotation technology brings AVL more efficient, precise, and accurate annotations. One of the most important uses of this data is to train ML models autonomous systems can learn from to learn how to navigate the world. In order to do so, we must annotate, fuse and label the data we have collected.
Annotation types supported:
- Bounding boxes
- Semantic segmentation
- Polylines
- Scenario labeling
- Keypoints
