Digital Map Vendors Are Jumping on the AI Bandwagon

On February 13, 2019, Japan-based Dynamic Map Platform (DMP) acquired Ushr, a Detroit-based digital map company, for US$181 million. The acquisition will create a single, enhanced platform for accelerated development and delivery of high-definition (HD) mapping solutions for autonomous vehicles and advanced driver-assistance systems. Ushr’s acquisition not only opens up the North American market for DMP, which is largely restricted to Japanese automakers, by adding existing business with GM, but also provides DMP with the much sought-after sensor fusion technology that enables the creation of high-definition (HD) maps.

Just a day later, HERE Technologies announced that it is investing US$28.3 million (or EU€25 million) into its Vienna-based international Institute for Advanced Research in Artificial Intelligence (IARAI). According to HERE, AI will be the technology developing next generation mapping services. This includes self-healing maps, location awareness capabilities in vehicles, and highly accurate optimization models for traffic, fleet management, and city infrastructure. HERE is expected to contribute its large repertoire of geospatial data to develop relevant AI models.

Investment into AI, especially by HERE Technologies, indicates the changing landscape of the industry. While existing business models that rely heavily on OEM and aftermarket remain lucrative, traditional digital map vendors can no longer ignore the influence and success of internet giants in location-based services for the consumer and enterprise markets. Free navigation services and location recommendation services from the likes of Google and other internet giants are disrupting the conventional business models of the industry. By getting into the AI game now, DMP and HERE are looking to maintain their competitive edge. By leveraging the latest development in machine learning, digital map vendors can kill two birds (i.e., further optimization of the digital map generation process and the integration of new capabilities to the digital map platform) with one stone.

At present, digital map vendors are relying on human employees to gather, prune, and update geospatial data, while incorporating aerial data collected from satellites and drones. Whenever there is an update to traffic conditions or road design, these updates need to be incorporated manually. As such, the entire process consumes large amounts of man-hours and resources, with little or zero real-time information incorporated into these updates. Many digital map startups have been looking at using various technologies to automate and make the process much more efficient.

One of them, DeepMap, is using Light Detection and Ranging (LiDAR) to construct digital maps. While there has been new advancement in the LiDAR technology that leads to the development of lower-cost LiDAR technologies such as micro-electro-mechanical systems (MEMS) and optical phased array (OPA), LiDAR technology has always been expensive. Other startups, such as lvl5, Momenta, and NetraDyne, are using camera arrays that come at a fraction of the cost of LiDAR. These red, green, blue (RGB) or time-of-flight (ToF) camera arrays, mounted on passenger cars and commercial vehicles, are powered by graphics processing units (GPU), such as NVIDIA Jetson, or application specific integrated circuits (ASIC), such as Intel’s Mobileye, to collect real time visual information about roads and streets. Aside from cameras, other sensors may include inertial measurement units (IMU), GPS modules, and radar. Deep learning algorithms are then used to compile the images to generate and update map information at near real-time in the cloud. The updates are then shared with the drivers. These processes are expected to be seamless and fully autonomous, requiring fewer human resources compared to the conventional method.

These solutions will make a huge difference in emerging markets. For markets with established and mature road infrastructure, the collection of digital map information can be less challenging, as the basic map information is readily available and fairly accurate. In addition, these markets will have established commercial telematics services, which provide a great foundation for map information. In large emerging markets where the infrastructure is less developed, such as India and Indonesia, map accuracy can be low and the adoption of telematics services may be limited. Logistics and delivery services may still largely rely on two-wheelers. The aforementioned low-cost method will allow digital map vendors to serve these cost-sensitive markets without incurring heavy capital expenditures.

In terms of the integration of new technologies, autonomous driving is obviously the key technology that digital maps and AI can enable. Three main components of autonomous driving, i.e. sensing, localization, and routing, require the combination of various sensors’ data to make decisions based on both the localization information and any new information coming from the sensor data that it might have to respond to in real-time. Internet giants such as Google and Baidu are at the forefront of this technology and, coincidentally, both companies have in-house digital map and AI capabilities.

In addition, an important secondary objective for map makers is the ability to leverage AI to create new and more accurate location-based services. Parking, traffic, etc. all require up-to-date maps and timely insight into the location and behavior of connected vehicles on the road. Investment in AI can enable map creators to scale both. For example, in-vehicle voice assistant has been very popular among digital map vendors. Internet giants such as Apple, Google, Alibaba, and Naver have enabled voice assistant in their own map platform for hands-free enquiry and navigation. SK Telecom, whose T-Map is very popular among road users in South Korea, has incorporated its telco voice AI assistant, NUGU, as a virtual assistant to perform searches within T-Map and provide point-of-interest (PoI) recommendations.

As these new use cases incorporate various data sources such as image, video, text, voice, and time-series data, the AI hardware deployed on the vehicles is likely going to be heterogenous in nature. A combination of central processing units (CPU), graphic processing units (GPU), digital signal processors (DSP), and image signal processors (ISP) will be required. Once processed at the edge, all this new information will be shared to the central cloud platform, where the master AI model is hosted. The master model will be trained and updated with new data inputs and the updated AI model shared with other vehicles under the same system. The complexity of these chipset technologies not only means that traditional digital map vendors may need to partner with chipset vendors, such as Intel, Qualcomm and NVIDIA, to offer end-to-end capabilities, but also favors companies that have made early investment in chipset capabilities, such as China-based NavInfo.