Overcoming the Challenges of 3D Mapping for Autonomous Cars

January 26, 2021

Story

For autonomous vehicles to progress to their next evolutionary phase, it is essential that vehicles have a more complete understanding of the environment in which they must safely move through.

In particular, to accurately and safely reach a destination, a vehicle needs to understand the available routes. Current technology uses markers placed along the road to help guide the vehicle.

While useful for many applications where routes are predetermined and fixed, such an approach has limitations. The vehicle has to follow a prescribed course and, if there is a problem such as a road closure, the vehicle has limited options for circumventing the problem with an alternative route. Perhaps most importantly, given the cost of deploying road-based sensors, the vehicle is limited in the number of destinations it can autonomously drive between.

Vehicles already have a number of cameras and various distance sensors such as those based on LiDAR. These cameras primarily provide Advanced Driver Assistance Systems (ADAS) with information about the objects located and potentially moving around a vehicle. This makes it possible for the ADAS to perform many tasks, such as alerting the driver if a person or another car is behind the vehicle when the driver is backing up.

Cameras on the vehicle, however, cannot provide all of the information a car needs to be able to drive autonomously in real-time. With access to a prepared 3D map of the environment, however, the vehicle can streamline route processing, support adaptive routing, and increase safety through better assessment of the movement of other vehicles. For example, the ADAS can have awareness of an upcoming intersection, whether it is heavily trafficked or not, and better predict how other vehicles may change lanes or turn with a minimal impact on real-time processing resources.

The Challenge of Building 3D Maps

One of the reasons 3D mapping is useful to autonomous vehicles is because it offloads the need for the vehicle to build its own map in real-time. Building 3D maps is extremely compute-intensive, and using pre-built 3D maps effectively offloads this task from the autonomous vehicle. When a 3D map is available, the vehicle can focus on identifying and tracking objects within the environment rather than having to first scope out the environment so it can then identify objects within the environment.

Building a 3D map requires that a mapping vehicle drive through the area to map and capture data critical for autonomous vehicles. The raw video data will then be processed into a format that is easier for the ADAS to utilize.

There are many challenges to designing a system that can capture the data needed to construct a 3D map:

• Real-time video capture is needed. The fastest PoE cameras on the market today require a 10GigE interface to store video without data loss.

• Multiple cameras must be captured simultaneously. The vehicle collecting data for 3D mapping has to collect video from multiple angles to capture an accurate representation of an area.

• High-performance computing platform required. The system has to complete many tasks in addition to capturing and storing video data, including 3D mapping model setup and SLAM processing.

• In-vehicle space is limited. The entire 3D map capture system has to be able to fit within a vehicle.

• Power must be optimized. The 3D mapping system must be optimized for in-vehicle operation. This includes keeping power requirements low as well as protecting the system from events such as power surges that occur when the vehicle ignition is active.

• Harsh operating environment requires rugged implementation. Automotive vehicles are a harsh operating environment. Factors such as temperature, vibration, and abrupt starts/stops can lead to unintentional disconnection of interfaces and boards that disrupt processing.

• The system must be easy to set up. Ease of setup is essential for simplifying vehicle operation, minimize opportunities for human error, and assuring reliable data capture.

3D Mapping: A Case Study

A Japanese software startup approached Vecow to build a 3D mapping system to support a wide range of autonomous vehicle applications, including self-driving cars, ADAS for passenger cars, golf carts, AMR, AGV, and drones. Vecow worked with the software company to build a custom real-time computing system that would provide the capabilities they needed to overcome the challenges associated with 3D map capture. The result is the ECX-2400 PEG + Dual PE-7004MX.

The ECX-2400 PEG (see Figure 1) is built on a workstation-grade 10^th Gen Intel® Xeon®/Core™ i9-10900E processor (CML) AI computing system with NVIDIA® Tesla®/Quadro®/GeForce® graphics. The i9-10900E uses the Intel W480E chipset with a 10-core CPU to support high-performance computing and multiple 5G/Wi-Fi 6/4G/3G/LTE/GPRS/UMTS interfaces for high-speed wireless data transfer. The 10-core processor easily handles 20 computing threads, supporting real-time live video capture from multiple sources at the same time as 3D mapping model setup and SLAM processing.

The PE-7004MX (see Figure 2) provides powerful interfacing capabilities to the ECX-2400 PEG with an Intel® X550-AT2 4-port X-coded M12 10GigE IEEE 802.3at PoE+ expansion card and PCI Express x8. It also provides up to 25.5W power output at 48V DC per port with PoE+ on/off control. Multiple 10G USB interfaces are available for high-speed data transfer to four front-access lockable SSD trays supporting RAID 0, 1, 5, and 10 for smart data protection. Two PE-7004MX can be used in tandem to increase the number of PoE+ ports available to support mapping systems with up to eight simultaneous video sources.

Together, the ECX-2400 PEG and PE-7004MX simplify video capture by requiring less system setup effort to increase performance reliability with no additional power connections. In addition, the system features flexible customized expansion to match varying project requirements, including dual PCIe x16 expansion slots, dual M.2 slots, and dual Mini PCIe for customized functions.

To fit within the space limitations of a vehicle, the system provides the most compact solution with eight X-coded M12 10GigE PoE ports on the market. This was possible through custom design by Vecow engineering experts. At the same design, the system was designed to be extremely rugged to assure reliability in the harsh operating environment of a vehicle. Each PE-7004MX features rugged X-coded M12 connections, a riser card for dual 10G PoE+ cards, a computer chassis with customized I/O, and optimized mechanical design for dual PCIe expansion cards. In addition, all components are rated for extended temperature operation (-25 to 60°C).

Finally, the entire system was built to optimize power consumption for in-vehicle computing. The ECX-2400 PEG serves DC 12V to 50V wide range power input with 80V surge protection. It also provides software ignition control to secure reliable system operation within the vehicle.

Solutions like the ECX-2400 PEG and PE-7004MX are critical to advancement of autonomous vehicle technology. With a rugged real-time AI computing system to build 3D maps in real-time, OEMs will be able to expand the capabilities of autonomous vehicles by freeing them from road-based markers and enabling them to safely and reliably travel all the roads of the world.