Inter-Vehicular Communication – From Edge Support to Vulnerable Road Users II

In this talk we first defined what is a Vulnerable Road User (VRU) from the point of view of many international standardisation entities. We then discussed four possible approaches for ITS to protect VRUs, highlighting pros and cons. First, the use of smart infrastructure with V2X capability. Then, cooperative perception by vehicles, followed by VRU-awareness messages sent by VRUs themselves. Finally, we discussed the use of edge/cloud support and introduced some open research questions that could be addressed during the rest of the seminar.

In this talk, I presented results from recent and newly-starting German and European research projects SecForCARs-SAVE, CONNECT, and ConnRAD where we investigate the role which trust models can play in securing complex cooperative, connected & automated mobility (CCAM).

CCAM systems are highly complex systems-of-systems (SoS) which are composed of many layers of subcomponents. Motivated by incidents like the log4j vulnerability, supply-chain security has recently taken up the challenge to evaluate security in such SoS. In our research, we investigate how to model trust dependencies in such SoS as trust networks or trust graphs in order to allow a quantifiable analysis of the effects that security incidents in one part of the system will have on other parts.

Based on earlier works, we were able to identify Subjective Logic and Subjective Trust Networks as a very useful formalism to model such trust graphs. In the talk, I illustrated this process on the example of Cooperative Intersection Management (CIM) and showed the steps that are needed for a MEC server to establish a trust opinion on the positions that vehicles send as part of their CAM messages.

In the following discussion, we elaborated on different aspects of such trust models, for example, the role of vehicle manufacturers as possible trust brokers as they constantly monitor their vehicle fleet through their backend systems and would be in a very good position to detect intrusions or other incidents that would reduce trust in a specific vehicle.

The first breakout session initiated with a discussion on the relationship between IoT and intelligent mobility applications, and identification of relevant applications to make discussion more concrete. A plethora of sensing and analytics applications have been considered in research related to mobility such as pedestrian flow detection, trajectory prediction, collision risk detection, user profiling, and so on. Considering vulnerable road users (VRUs), sensing and analytics data services would make use of data such as video data, GPS trajectories, and vehicular sensing.

The sensing may have two types of goals: real-time traffic management and safety applications, or feeding urban planning. Analyzing such data in large geographical domains would bring data communication challenges as well as challenges in the computing and Artificial Intelligence (AI), where specialized algorithms in distributed analytics would be studied. Machine Learning (ML) training and inference problems can be considered for developing vehicle-specific (in-vehicle), local, cooperative and federated models. Particularly, federated and split learning are an on-going research directions that address the distribution of the more computationally expensive phase of ML models: training. These two types of algorithms address privacy constraints by avoiding centralization of the raw data. Other on-going efforts in ML that are relevant to low latency include early-exit models, which process data through the whole pipeline only when necessary, e.g. for increased confidence. This would allow light local processing, moving data out of the mobile device only for some specific (detectable) situations. The distribution of ML model training involves several challenges related to the data itself. Besides addressing the need for i.i.d samples or the bias caused by non-i.i.d samples for most models, some areas currently under-explored are the distributed data pre-processing (e.g. local statistical measures may differ from global ones), and how the annotation of the data in a distributed setting could be achieved.

The trade-offs between different degrees and type of distribution, convergence speed, networking costs and model accuracy are yet in the realm of research. Modelling such trade-offs was identified as a valuable research direction. Distributed computing should make use of cloud and edge resources efficiently, such that the quality-of-service requirements, e.g. latency, from both the communication (network latency) and computation (virtualization and AI latency) angles would be satisfied. One may consider the networking as “in-vehicle” where wired communication would be utilized whereas the information that goes out of the vehicle should be transferred through wireless communication. For the computing side, edge computing may be applied in-vehicle, road-side, or edge data centers that are in the vicinity, materializing the vehicular edge computing paradigm. A brief call of attention was made to the different semantics of the fog, edge and cloud nomenclature in different communities, namely the Internet of Things.
Distributed sensing data collection and analytics are of utmost importance for improved safety applications for cars, bikes, and pedestrians. For instance, cooperative sensing and perception has the potential to greatly improve the VRU safety in various areas of traffic in and out of cities such as traffic intersections, pedestrian crossings, or blind spots. On the other hand, there is a challenging decision making process between data offloading for improving application performance and keeping critical data in-vehicle for privacy and liability reasons. Especially, liability and accountability concerns are likely to play a determinant role on these scenarios. Other than the safety, there can be various mobility applications. A few of these applications, which can be enabled by distributed sensing and analytics, are listed below.

• $\mathrm{■}$

  Digitalization of the cities: e.g. for improving navigation, transportation and parking services

• $\mathrm{■}$

  Pothole or obstacle detection: Enabling comfort and efficiency.

• $\mathrm{■}$

  Dynamic infrastructure: making dynamic changes to the infrastructure for cost and energy efficiency, e.g. pop-up bike lanes.

The above-mentioned sensing and analytics applications mostly involve dealing with personal and sometimes confidential information, thus privacy-aware system design is a key aspect. The privacy-aware design would have particularly more importance when the new systems evolve from research-level prototypes towards real deployments. Trade-offs between utility and privacy are not well understood, and privacy by design should become the state-of-the-art. Yet, a quantification of the trade-offs would be valuable.

The future of Cooperative-Intelligent Transportation Systems (C-ITS) and the challenges for integration of VRU into the picture was widely discussed. It was consensual that to a large extent the technical problems have been covered in previous research, and the biggest hurdles to actual implementation are of societal, legal, political and economic nature. Nevertheless, several open research directions were identified.

• $\mathrm{■}$

  Improved large scale simulation models and digital twins enabled by high performance computing will enable a better understanding of behaviours. Current simulation solutions are often closed, proprietary, expensive, and of limited access. Lowering the cost for such solutions, e.g. using open source to facilitate evolution, expansion and integration of different simulation environments, enabling this integration on multiple-locations, etc would be of great value to a broad research community. Current status is assessed as very early infancy.

• $\mathrm{■}$

  Sensing and analytics plays a key role to build these models: mobility micro-models for traffic participants, especially VRUs, behavioural models (operational, tactical and strategical) in complex and safety relevant situations, traffic light models. It is of special relevance to study and model unexpected behaviours or behaviour/ intention change, as these are more likely to cause hazardous traffic situations. The metrics to validate such models are also in their infancy.

• $\mathrm{■}$

  Little data exists about accidents and their analysis, and access to existing data is a challenge for collaborative research, e.g. accident analysis databases like GIDA require very strict NDAs. User interface research for interaction with VRUs is another research direction with significant gaps.

Much of this research is strongly inter-disciplinary, and disciplines like transportation, urban planning and human factors need to be involved.

The breakout discussion on edge computing started with the discussion for relevance and applicability of edge computing in terms of the communication and cost of resource usages. There are several questions raised for the applicability of the edge computing:

• $\mathrm{■}$

  Who should manage the edge servers?

• $\mathrm{■}$

  Which computations should happen at the edge?

• $\mathrm{■}$

  What are the communication requirements of mobility applications?

Starting with the first question, there is a question on whether mobile service providers or original equipment manufacturers (OEMs) should deploy and host the edge servers. For instance, mobile service providers may not have interest to realize dense deployments but OEMs can build edge servers on the cars and loads of data can start coming from the cars. For the second question, high-definition map building through video could be considered. Such application would cause high demand in terms of computation cost, bandwidth, and latency due to transfer of videos and the tasks of high-definition map building. There are, however, many more lightweight applications, which intelligent transportation systems applications could benefit from.

As a benefit of edge computing, the edge layer can serve as a buffer, where various computation tasks such as preprocessing could be performed on the edge; and extracted information could be shared with the cloud. Considering the other way around, the edge servers might share data with the vehicle for the internal computation and actuation of the vehicle. However, such critical scenarios will need to be carefully designed to avoid security vulnerabilities. For instance, steering a vehicle by bringing data from edge or cloud might create vulnerabilities to attacks. Furthermore, as vehicles move, they will occasionally be in out-of-coverage areas of the service providers. Thus, basic services and decisions such as steering can stay in the vehicle itself, whereas certain information that could not be collected through in-vehicle sensors may enable a smoother ride.

Edge computing would enable various applications, some of which are discussed during the breakout meeting. These applications include “cooperative” applications that require multiple vehicles as opposed to having a single vehicle behaving independent from other vehicles. At the initial phases of edge computing, a pragmatic approach would be to start with a “minimum viable edge”. The minimum viable edge would have services that are beneficial and easily applicable. For instance, assisting consensus building or creating local dynamic maps based on information collected from various vehicles could be implemented on the edge. For the latter, information can include traffic congestion, obstacles (e.g., potholes), disasters (e.g., water pipe burst), and other unexpected events (e.g., animal nearby). Moreover, applications that are not safety critical such as parking services could be more efficient through the application of edge computing.

One important aspect is the physical placement of the edge computing – if realized on cars. Several possible options exist for the physical placement:

• $\mathrm{■}$

  Intersections, where many vehicles pass by regularly

• $\mathrm{■}$

  Parked vehicles, where the computing platform could be made available

• $\mathrm{■}$

  Charging stations, where vehicles wait for relative long times.

The final part of the breakout discussion focused on the business models for future mobility use cases in terms of edge computing. Recently, vehicle manufacturers have become more like “data” companies as they collect mobility data from the vehicles and users. For the data ecosystem aspects, developments in the fields data contextualization (e.g., FIWARE, smart data models), for understanding data and creating value, as well as data spaces (e.g., Gaia-X, IDSA) for data sharing and exchanges between different parties are becoming highly relevant.

Trust can be considered a key aspect of resilient systems, reliably assessing a systems trustworthiness enables informed decisions with respect to, for example, safety- or security-critical functions. This working group discussed trust in cooperative, connected & automated mobility (CCAM) starting with looking at different dimensions of trust. We distinguished a technical, human, and regulatory notion of trust. The technical perspective is based on a notion of functional trust, i.e., the trust that one entity puts in another entity to perform a specific function in a trustworthy way. Alternatively, this can also mean that named other entity can provide certain data accurately. For example, this can refer to another vehicle correctly reporting its position in a CAM message where a receiving node has to rely on this data to predict collision risks. Modeling such trust relationships between components in an automotive system of systems leads to a trust graph that could be modelled using a formal logic to quantify the amount of trust and reason over trust relationships. Subjective logic trust networks are one suitable formalism which was illustrated in the plenary talk of Frank Kargl. With such an approach, a functional model of a system could be augmented with a trust model that allows to reason about technical trust in the system both at design and at run-time. Subjective logic provides the appealing property to allow reasoning under uncertainty with incomplete evidence. Open questions here include how and where to find the initial evidence for trustworthiness to feed the trust model with concrete data. This could come from looking at trust- or node-related trust as distinguished in a survey on misbehavior detection [1]. Furthermore, the structure of trust networks and reasoning approaches and the expected and required levels of trust require further investigation.

Bicycles are a healthy and environmentally friendly transportation mode that is increasingly used for commuting. Connecting bicycles to other vehicles is an enabler for a large variety of services, from safety to infotainment. Conversely useful and comfortable applications could make cycling more attractive. This talk shows demonstrations of motivating applications supported by Bluetooth Low Energy (BLE): a stolen bike detection system [1], and a broadcast walkie-talkie for a platoon [2].

The talk evolves to explore connectivity aspects needed to support such applications, both to connect bicyles to one-another and to the infrastructure. A characterisation 2.4 GHz operating technologies on bike-to-bike links using commodity hardware shows that BLE range exceeds 50m at low packet loss rates [3]. A dependence on relative bike positions was identified. Anechoic chamber measurements with bikes allowed to characterise the dependence of bike-to-anything links on antenna position and bike material [4].