Measurements show the mean (solid line) and standard deviation (shaded area) of three runs. b) Waste of time; (c) Number of collisions. Average accuracies obtained for each of the tested algorithms and each of the proposed accuracies.
Introduction
- Thesis motivation
- Thesis objectives
- Ph. D. dissertation structure
- Thesis contributions
Assess the state of the art of the topics covered in this dissertation: ITS, AI, UAVs and wireless communication technologies. Detailed study of the operating principle of the most important adaptive traffic light control systems.
Conceptual background
Introduction
Intelligent Transportation Systems
The main application of BLE in ITS is to provide low power consumption and low implementation cost communication for sensors used in ITS that provide real-time information on several environmental parameters such as the number of vehicles, vehicle direction, road conditions, pollution, noise, etc. It will also support a wide range of human-oriented services, such as cooperative autonomous robots, the metaverse, and real-time services based on cloud and virtualization.
Artificial intelligence
- Machine Learning
- Deep learning
- Reinforcement learning
- Deep Reinforcement learning
- The emergence of AI in ITS
In model-based algorithms, the RL method uses a model of the environment (dynamics), i.e. transition function (and reward function) to estimate the optimal policy. Models can be given (e.g., a game of chess, where the valid moves and the probability that the opponent will make a move are known) or they can be learned during training.
Conclusion to this chapter
Here, artificial intelligence must be combined with wireless communications, sensors, computing and processing power to achieve a system that is able to adapt to a changing and unpredictable environment, learn and adapt in real time, operate autonomously and proactively, and be able to communicate with other actors to achieve deep collaborative and cognitive intelligence.
Unmanned Aerial Vehicles
Introduction
- Mobility models
- Positioning protocols
- Propagation models
- Routing protocols
- Experimental setup
- Results
- Conclusions
Another approach to FANET deployment is based on the capabilities/requirements of the network to be deployed. Radiation pattern of isolated WiTi on the U2U link @ 5 GHz using the antenna configuration shown in Figure 3-13: (a) X-Plane; (by plane; (c) Z-Plane. UAV onboard WiTi radiation pattern for the antenna configuration shown in Figure 3-13 on the U2U @ 5 GHz link: (a) X-Plane; (by plane (c) Z-Airplane.
A Comparative Performance Evaluation of Routing Protocols for Flying
- The importance of real experimental studies
- Experimental setup
- Results
- Conclusions
The results showed that the impact is mainly due to the carbon fiber chassis, with the impact from engines and propellers being negligible. 263], a routing protocol based on the geographical position of the nodes, called Ground Control System-Routing (GCS-R), was developed. On the other hand, we can observe a remarkable difference in the performance of the results between the 2.4GHz band and the 5GHz band.
Conclusions to this chapter
Routing protocols are particularly important, especially in large networks, where the amount of data sent is large and therefore very important for the proper functioning of services. Therefore, the choice of the appropriate protocol depends on several factors, such as the context, the application and the requirements of the users of the service. The results showed that it is crucial to choose a routing protocol before setting up a network in order to achieve optimal performance of the entire network.
Publications associated with this research
Finally, the last section addresses the question of how the different routing protocols affect the service FANETs offer. Therefore, in the second part of this section, we tested and compared the performance that three proactive routing protocols for FANETs can offer and present the results in terms of throughput and packet loss.
Smart Cities and Pedestrians
Introduction
In addition, this type of technique presents a higher computational processing cost, both during the training of the detection systems and in the inference. Active methods are those that require the action/response of the mobile device to be counted and passive methods are those that obtain an estimate of the number of people only by analyzing the Probe Request messages sent periodically by the mobile devices to be counted. This randomization policy depends on the manufacturer of the communication chip and the condition and characteristics of the mobile device.
Related works
- Device-Free approach
- Device-Based approach
An overview of the fields in the Probe Request message can be found in Figure 4-2. As you can see, within the Probe Request field, the address of the device sending the message, i.e. the MAC address, is sent unencrypted. In this work, they used Probe Request message capture to locate people in disaster/emergency scenarios.
A WiFi-based method to count and locate pedestrians in urban traffic
- Introduction
- Experimental setup
- Our proposal
- Experiments & Results
- Conclusions
In addition, the influence of the frequency of sending Probe Request messages on the accuracy of the proposed algorithm was studied. The third set of experiments was conducted to obtain the overall accuracy of the proposed algorithm. The last set of experiments consisted in analyzing the influence of the sending period of the probe request message on the accuracy of the proposed algorithm.
Counting and locating people in outdoor environments: a comparative
- Introduction
- Experimental setup
- Results
- Conclusions
A comparison of ML algorithms against the proposed algorithm was performed on the discrimination of pedestrian behavior (moving or static). The input data for these ML algorithms was the same as that used by our proposal, i.e., the power variance of polling request messages captured across all DAUs. In addition, the average accuracy values obtained for each of the algorithms are presented in Table 4-8.
Pedestrian Characterization in Urban Environments Combining WiFi
- Introduction
- Experimental setup
- Experiments & Results
- Conclusions
How each of the ML algorithms studied works: (a) Logistic regressor; (b) Gaussian naive Bayes; (c) support vector machine; (d) k-nearest neighbor; (e) Random forest. A confusion matrix was used to compare the performance of the different ML algorithms studied. The inference results on the test data set after training can be seen in Table 4-16.
Conclusions to this chapter
It is important to keep in mind that our algorithm requires simplicity and lightness so that it can be executed in real time on low-cost, low-power embedded devices. The overall performance of iPCW was a discrimination accuracy of more than 98%, as well as a positioning accuracy of more than 92%. In this way, an advanced control policy can be obtained that can monitor both vehicular and pedestrian traffic, maximizing road safety and minimizing waiting times.
Publications associated with this research
After the selection of RF, a new algorithm was designed called Intelligent Pedestrian Characterization Using WiFi (iPCW). In conclusion, the systems proposed in this section can provide traffic control systems with the ability to involve pedestrians in their control decisions.
Smart Traffic Light Control – AI approach
Introduction
State of the art
In the second group we find the work of Sánchez-Medina et al. Also, in [342] we can find an approach that unifies the communication between track and GA devices. If we focus our search for papers that use AE, we can find papers that use Ant Colony Optimization (ACO) processes.
Study on the influence of traffic signal duty cycle duration at a single
- Introduction
- Experimental setup
- Results
- Conclusions
In this figure, the behavior of each of the analyzed metrics can be seen as a function of the duration of the green time (between 15 s and 100 s), which directly affects the cycle time of the traffic lights. Finally, Figure 5-3 f shows the optimal green time value for each of the analyzed metrics, as a function of the incoming vehicle flow per lane. Shown in each of the subfigures (a-e) are the different metrics analyzed as a function of the green time of the traffic light cycle.
Intelligent IoT systems for traffic management: a practical application 110
- Genetic Algorithms
- Experimental setup
- Results
- Conclusions
A representation of the simulated vehicle flow in the training scenario can be seen in Figure 5-6. The evolution of the fitness level (average waiting time per intersection) reached by iREDVD during the generations of the genetic algorithm can be seen in Figure 5-11. This behavior can be seen in Figure 5-12a and Figure 5-12b, where it is highlighted how iREDVD proactively adjusted the traffic light cycle time as a function of the simulated traffic flow, both in the training scenario (Figure 5-12 a). and in the test scenario (Figure 5-12 b) in relation to unknown situations.
Conclusions to this chapter
Publications associated with this research
Guillén-Pérez, A.; Cano, M.-D., “Optimización de un sistema inteligente de control y gestión del transporte en intersecciones mediante un algoritmo genético”, V Jornadas de Doctorado UPCT, Universidad Politécnica de Cartagena.
Interoperability of Connected Autonomous Vehicles and
Introduction
- Autonomous Intersection Management
Under this approach, only one vehicle was allowed to enter the intersection at a time, regardless of the route the vehicles took. The second approach only considered the conflict points of the routes the vehicles followed. TD3 updates the actor policy (and the target networks) less often than that of the critics.
State of the art
Curriculum learning [389] consists of sequentially training an intelligent agent on a task that gradually increases the complexity of the task to be performed. For example, in object manipulation and classification tasks, different objects can be added when the task results are satisfactory and stable results are achieved. The results they obtained showed the superior performance of their proposal in three out of four scenarios tested against FCFS and semaphore-based control algorithms.
RAIM
- Introduction
- RAIM - Reinforced Autonomous Intersection Management
- Experimental setup
- Results
- Conclusions
Additionally, a graphical representation of the neural network architecture designed for RAIM can be seen in Figure 6-3. For the optimization of the controller modeled by the neural network seen above, the reinforcement learning algorithm TD3 [388] was used. The results obtained showed excellent stability due to the set of algorithms used (PER and curricular learning).
Multi-Agent Deep Reinforcement Learning to Manage Connected
- Introduction
- adv.RAIM – advanced Reinforced Autonomous Intersection
- Experimental setup
- Results
- Conclusions
Finally, adv.RAIM was compared with more recently published algorithms, such as an intelligent traffic light control system (iREDVD) and a previously proposed AIM. Following the methodology used in the previous work, adv. RAIM was trained by following the curriculum through Vetë-Loja. More interesting is if the performance obtained by adv.RAIM is compared with those of the AIM algorithm proposed by Qian et al.
Learning from Oracle Demonstrations – A new approach to develop
- Introduction
- Imitation Learning
- Learning from Demonstration
- TD3 from Oracle Demonstration – TD3fOD
- Experimental setup
- Results
- Conclusions
These demonstrations are used to train a new agent by supervised learning, learning to "mimic" the e pert agent's behavior. As can be seen in that equation, the influence of the Oracle on the TD3 actor is controlled by the parameter 𝜏1. This start can be seen in the change in the trend of the measurements that were analyzed at the beginning of this phase.
AIM5LA: A Latency-Aware Deep Reinforcement Learning-Based
- Introduction
- AIM5LA
- Experimental Setup
- Results
- Conclusions
- Introduction
- Key Enabling Technologies
- Connected Autonomous Vehicles Framework over 6G
- Use Case
- Conclusions
This network handles encoding the previous delays of other vehicles in order to take these behaviors into account when controlling the ego vehicle. By means of this message exchange protocol, AIM5LA can calculate the delay that exists in the communication channel with each of the CAVs. This timestamp is the internal IM clock before the message is sent.
Conclusions to this chapter
Publications associated with this research
General Conclusions and Future work
Introduction
Key future research directions