3.1. Autonomous Vehicles Detection Capability at Intersection and for Horizontal Road Markings

The quality of horizontal road markings plays a crucial role for autonomous vehicle detection, not only to avoid conflicts but also to facilitate pedestrian individuation, for instance, in zebra crossing points. Hence, ADAS (Advanced Driving Assistance Systems) equipped by AVs must ensure an efficient and smart detection of road markings, primarily due to their frequent deterioration (mainly caused by severe weather and actions from rolling wheels) and also to poor maintenance. Road markings are fundamental for vehicle lane-keeping, with straight or curved lines substantially characterizing these signal types; additionally, horizontal markings may also include directional arrows, helpful for the AVs’ decision-making process. However, directional arrows can be perceived differently than simple lines by visual cameras due to several factors like angles of view, line thickness or interference of shadows projected by nearby objects. In AVs, the detection process of horizontal road markings occurs through vision-based systems that result to be the more appropriate, mainly for economic reasons thanks to the inexpensiveness of vision cameras. Moreover, these systems also allow the creation of 3D input data whose quality is optimal as one of LiDAR’s systems [25], useful considering that visual data is the most suitable input for human drivers.

Therefore, road marking detection has a dual approach, the first related to AV detection capabilities and the second to the quality of horizontal road markings. The latter represents an aspect of vehicle circulation affecting safety regardless of autonomous performances and MV (machine vision) capabilities. Since 1992, Miller highlighted how to correct maintenance of road markings involves at least sixty times less than costs caused by incidents caused by a lousy perception of roadway spaces. The parameter that mainly affects MV’s road markings perception is the retro-reflectivity (RL), especially in low light hours. Such feature represents the capability to reflect radiations and can be measured as mcd/m²/lx. In 2016, Matowicki et al. demonstrated that, when road markings had an RL = 50 mcd/m²/lx, MV systems began to show malfunctions, while other studies, like the one of Davies in 2017, reported that MV was also affected by width and color, where white was more suitable to be detected compared to yellow [2].

Several methods have been studied to facilitate road markings detection with various processes. For example, Hechri and Adellatif provided a system characterized by two video-based driving tasks. At the same time, Lee combined AVM camera features with LiDAR’s data. Ultimately, Zhen Kang and Qiao Zhang proposed an algorithm able to speed up map rendering by solving point cloud data [26]. Some of these studies were also analyzed through a forecasting process with simulation software. Le-Anh Tran and My-Ha Le carried out their proposal of a robust road lane markings detection system with simulator CARLA that showed how the system resulted inefficient in straight parts. At the same time, it did not present beneficial results in curves [27].

It is possible to state that the quantity of horizontal markings is related to the roadways context dynamism, meaning that, whenever the possibility of change directions and actors being part of a specific context is elevated, the high quantity of roadway markings will characterize the environment. Therefore, intersections represent a critical issue for marking detection, but for instance, a vital detection challenge for AV sensors is the work area. Work zones are subject to unexpected continuous variations, mainly due to vehicle movement. That is why horizontal markings have a crucial role in implementing a safe work context, especially in highways or freeways where speed ranges are elevated. For these reasons, some studies considered proposing new pavement markings for work zones, as reported in Table 3 and, in some cases, with no predetermined itinerary vehicle so with driver presence. Generally, in a work zone, a significant trial for autonomous vehicles is the reduced lane width between 3.5 and 2 m, if not even less. Different types of barriers or furniture can also be positioned in such lanes that the AV detected as obstacles can cause vehicle sudden stops when is not required. Ultimately, one last critical aspect consists of using various pavement marking materials to cause a high quantity of data collection for AVs when required just the same function or markings detection. Then it would be helpful to the deployment of uniform materials [28]. One aspect that can help solve the aforementioned problems is the lower speed in work zones that allow detection sensors to get better environment perceptions. Therefore, speed range is fundamental [29]. Highways or freeways can be subjected to many conflicts if signals and markings are not efficiently detectable. It has to be highlighted that autonomous vehicles’ processes related to environment detection can co-occur to a data management and decision-making process of multiple vehicles handled by a central infrastructural system. Then, it is helpful to apply methods that facilitate the operation previously cited. For example, it is possible to apply protocols related to vehicle intersections behavior based on V2V communication. Some of these protocols are Advanced Maximum Progression Intersection Protocol (AMP-IP) and Maximum Progression Intersection Protocol (MP-IP). MP-IP objective is intersection throughput increase concerning safe crossings until its primary goal of passage safety is satisfied all vehicles are allowed to cross the intersection, also in case of potential conflicts. AMP-IP considers a general conflict cell, allows low-priority cars to cross the intersection before high-priority vehicles arrive [30]. Automated Intersection Management or Intersection Managers (IM) are systems designed to interact with vehicles approaching intersections, aiming to improve traffic safety and efficiency.

Table 3. New markings for AVs movement in work areas [28].

Marking Type	Pavement Marking
Lane Open
Lane Closed
Right Turn Single Lane Ramp Closure
Right Turn For Single Lane Right Ramp Closure
Merge Right
Left Turn Two-Lane Ramp-Left Lane Closure
Left Turn Two-Lane Ramp-Right Lane Closure
Left Turn Single Lane Ramp Closure
U-Turn
Start Work Zone
End Work Zone
Work Zone Edge Line (Contrast Pavement Marking)

One of these systems’ benefits is that they can reduce the probability of accidents caused by human mistakes, often due to improper driving evaluations or impulsive driving behaviors, due to emotional decisions or instinctive actions rather than logical analysis [31]. Driving reaction times during emergency maneuvers are very short and often will not leave the driver enough time to decide the next move wisely. Automated Intersection Management (IM) systems work by exploiting the communication between vehicles and devices placed on intersections. Vehicles approaching the intersection at constant speed transmit a speed query to the IM, receiving a Yes/No reply depending on other simultaneous requests from other vehicles to optimize flow and manage safety. In case of a positive reply, the incoming vehicle is authorized to proceed and can drive forward, crossing the intersection at the requested constant speed. Should the answer be negative, in that case, the vehicle will slow down and resend new queries to the IM. Once the vehicle crosses the intersection, another signal is sent to the receiving device, communicating that the maneuver has been completed. This system increases the admissible road traffic volume and relies on a large and increasing amount of data to be exchanged between AVs and IM. Hence, its efficiency depends on the accuracy of this shared data. Robust Intersection Manager’s system adds the benefit of tracking vehicles’ trajectories approaching an intersection associated with a Velocity of Arrival (VOA) and Trajectory of Arrival (TOA) as they get close to the junction. This situation leads to increased accuracy and higher safety levels. This system is still subject to hardware or communication errors and even to hacking attempts interfering with vehicle positioning and speed detection. Figure 4 below depicts the dynamics of an accident happening with these systems.

Figure 4. Accident dynamics on IM systems [31].

The use of new strategies, based, for instance, on the Right-of-Way Assignment (RWA), can improve traffic dynamics in the presence of non-signalized intersections. This strategy is based on marking short-term decisions for the right of ways and on one-to-one communications. Other assumptions include that two vehicles can cross the intersection simultaneously and that pilot vehicles can obtain information about speed, location, and driving intentions of other surrounding vehicles. Moreover, priority levels and precedence rules are defined. A related strategy, called Cooperative Driving Strategy (CDS), introduces the assumption and possibility that the information connecting all the elements involved in traffic (pedestrians, bicycles, and vehicles, to name a few) is available via a V2X communication system, which can cover up an area of 300 m or even more. This methodology splits the junction in a grid made of several smaller areas, regulating vehicle traffic by setting up the minimum time interval between two vehicles. Results from dedicated research studies [32] show that this approach leads to higher performances compared, for example, to a traditional Responsibility Sensitive Safety (RSS) approach (which applies a typical First-In/First-Out (FIFO) strategy with few communications). Figure 5 and Figure 6 show a significant difference in performances between RWA/CDS and RSS, specifically in terms of average delay/arrival date and throughput/arrival rate ratios.

Figure 5. Comparison between RWA, CDS, and RSS in terms of average delay/arrival rate [32].

Figure 6. Comparison between RWA, CDS, and RSS in terms of throughput/arrival rate [32].

Simulation models for AVs at intersections can also be trained thanks to machine learning, specifically Reinforcement Learning (RL) and computer vision. Aerial photographs can provide visual information about the road intersection to create a simulated model. The vehicles’ position, both autonomous and manned, can be detected using deep neural networks of various kinds [33]. Vehicles in the intersection can be detected with the segmentation technique, starting from a full-color aerial image of the crossroad captured with drones and then post-processed, moving through computerized Fully Convolutional Network (FCN), creating a B/W image which then becomes clustered through a DBSCAN. With the OpenCV CVAT software tool, vehicles on the aerial images are marked and labeled with polygons (rectangles and ellipses), indicating the relative location. Possible trajectories are marked with polylines as well. Vehicles are finally manually removed from the image, leaving the polygons on. The image becomes then binarized and clustered into closed regions with the DBSCAN algorithm, filtering and discarding the smaller regions. From this point onwards, the simulation is undertaken thanks to the application of complex neural network algorithms that eventually lead to a calibrated and portable model, which can then be included in larger Automated Intersections systems.

Like RL and computer vision, Neural networks Systems (and generally speaking systems based on machine learning) often require advanced and expensive hardware tools like for example survey drones with cameras and powerful computers, which are not always affordable for common application. However, they can be seen as an example of how current technology for AVs has become sophisticated, considering how drones and surveying sensors have quickly become mainstream in recent years. In addition, there is always the risk of malfunctions or even hacking attempts to cause some damage. In addition to technical constraints, which can affect accuracy and influence costs, the use of drones might require special authorizations and licenses, which is another parameter to consider. Negligence, in this sense, might lead, for example, to fines, trials, or any other legal inconvenience with reputational and financial consequences. On the other hand, the achieved benefits are related to traffic optimization and its subsequent effects, the portability and flexibility of the simulation models, and the improvement of the modeling accuracy. The flowchart in Figure 7 summarizes the process adopted in this method.

Figure 7. Flowchart of the RL: computer vision process for an intersection [33].

It is helpful to remind, with a brief excursus, that artificial or electronic neural networks are computational systems inspired by and resemble biological neural networks, i.e., the complex system of neurons (nodes) and synapses (connectors) typical of animal brains. These systems allow the reconstruction of real-life scenarios characterized by the presence of scattered and interacting elements. It is beneficial, for instance, in camera detection whenever the nodes are hard to spot due to bad weather conditions, low light conditions, or obstacles concealing them [34]. Image segmentation can be performed with systems like Structured Inference networks [35], based on the assumption that the analyzed data conforms to a Gaussian State Space Model (GSSM). Commonly, algorithms to train neural networks fall into three major types: Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient [36]; adopting the best fitting algorithm depends indeed on the geometry of the collected dataset and the surrounding scenario. Deepening into Artificial Intelligence and Machine Learning would require a reliable dissertation as the subject is pervasive and has its particular field of study. Therefore, it will not be treated in detail in this article. Other similar approaches allow modeling vehicle trajectories and their variations for vehicles turning at intersections [37]. One of the methodologies adopted for this purpose exploits the principle of “Minimum Jerk”, i.e., a model initially developed to describe a human arm’s biodynamics in a plane, later extended to robotics eventually to AVs for studies on motion control. This model states that a jerk function can represent movements performed by the arm to carry out tasks like writing or drawing. The minimum jerk is found by minimizing this function through integration within an interval between a given initial location and a final location within a given time. The Minimum Jerk model and other support models aimed to identify minimum speeds and the vehicle-related position [38,39]. When inserted into a simulated framework, these allow estimating turning vehicles’ trajectories at an intersection. These estimations can be incorporated into automatic intersections to achieve the benefits already mentioned in this paper. It is interesting to remind that simulations created with all these models must also deal with human driving behaviors, which statistically depends on parameters [40], such as age, driving skills and training, and state of lucidity and attention, not to mention all those random parameters that could be present (even the presence of an insect could disturb the driver and cause an accident). The implementation of AVs [41] would ideally fix these issues, as the model’s random or less predictable components are removed since the driver is absent. This situation should lead to even higher optimization levels, increasing safety and reducing travel times, to mention a couple of benefits. Figure 8 summarizes with a flowchart the process and steps that define the Minimum Jerk approach. In contrast, Figure 9 shows an example of the computer output, reporting the comparison between paths (a), speed profiles (b), acceleration profiles (c), aggregated speed profiles (d), and aggregated acceleration profiles (e) for different exit speeds.

Figure 8. Flowchart of the process for Minimum Jerk model [37].

Figure 9. Computer output from a Minimum Jerk model—comparison between several traffic parameters for various exit speeds [37].

3.2. AVs Effects on Roadway Pavement

Once treated themes related to AVs’ advantages and expected consequences, it is necessary to focus on infrastructural design features [42,43]. One aspect can regard the change of load layout applied on road pavement. As previously argued, AVs would be able to keep with high precision in lane position. It is possible to consider that each vehicle’s wheels would stress road pavement always in the same areas and with almost constant speed in a context aimed at AVs exclusive transition. Ultimately, it is expected that road pavement would be stressed by vehicle loads not homogeneously, like for vehicles driven by a human driver (since each driver has its way of driving), but in a cross-section area that could be almost constant [44,45]. Table 4 reports an overview of AVs’ effects on roadway pavement.

With the introduction of autonomous vehicles, ordinary traffic will be less subjected to vehicles’ wandering phenomena than channelized traffic. Such conversion can have several consequences, such as greater rutting, significantly increasing longitudinal fatigue cracks. In this case, heavy vehicles would have to be mainly analyzed because tires could cause an elastic deflection bowl. This phenomenon is attributed to the possibility that AVs provide to reduce heavy vehicle headways. It was demonstrated by De Beer (1992) that, considering loads cycle application applied quicker, due to high frequency related to headway reduction, road pavement keep a slightly deflected condition that is, in turn, stressed again by following load axis application.

So, the subsequent deflection will be higher. Such issues can be managed with a technique named “strip road concept,” consisting of installing two parallel strips made of asphalt, concrete, or improved soil positioned in correspondence to specific vehicles’ wheels that will transit on that roadway. Significant localized stress pavement is subjected to AVs. It was detected during accelerated pavement testing (APT). It is noted, in Figure 10, different pavement responses under expected AVs and conventional vehicle loading respectively channelized and wandering traffic conditions standard hot mix asphalt surfaced.

Figure 10. Channelized versus wandering permanent deformation response [46].