Sensor fusion for cooperative perception in intelligent transport systems

• Autor(a) principal: Pinho, Hugo Emanuel Silva
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/40214
• topic: CPM; Connected vehicles; Sensor fusion; Autonomous vehicles; RSU; ITS; C-ITS; Vehicular communications; Collective perception; Image classification; YOLO

In a world that is becoming increasingly interconnected, autonomous driving systems necessitate sturdy communication platforms. Modern vehicles operate as machines and data centers that gather and distribute data from various sensors and devices to improve their cognitive processes. Accuracy, reliability, and adherence to communication protocols are crucial as cooperative systems increasingly rely on this information. Consider a scenario where an autonomous vehicle approaches an intersection and receives a message from another vehicle, indicating that it is slowing down to allow for the safe passage of the autonomous vehicle. Regrettably, due to incorrect information in the message, the autonomous vehicle assumes that the other vehicle has stopped and proceeds without reducing its speed, or in the same scenario, the infrastructure sends misleading information to the vehicle indicating a green traffic light at the intersection, while it is actually red. As a result, the autonomous vehicle crosses the intersection unsafely, disregarding the correct signal. These situations demonstrate the critical importance of data quality in cooperative perception and connected vehicle systems. Incorrect data, inaccuracies, or unsafe can lead to faulty decision-making by connected vehicles and infrastructures, endangering the safety of road users. The primary objective of this study is to overcome the challenges of object classification in cooperative perception messages transmitted by roadside units to ensure higher accuracy and reliability. Radars installed on roadside infrastructure accurately detect object positions and speed, but face difficulties in determining the type of object detected. This dissertation relies on image classification methods to address these challenges and to offer contextual information. This information is later integrated with the detected object data, utilizing an event-based sensor fusion model. This model incorporates the latest artificial intelligence algorithms and late fusion techniques to fuse data from various sensors. Late fusion was selected because it enables the utilization and integration of individually categorized data into a unified service. The study’s findings showcase the methodology’s effectiveness at diverse times of the day, which historically hindered the performance of roadside infrastructure systems, like camera systems. The fusion model significantly improved data classification during nighttime, radar comparison with fused data showed an improvement of 9.85%, while during the daytime, it increased by 23.56%. This study presents an effective approach to enhancing a radar’s comprehension of its surroundings, particularly when it lacks precise image classification systems. By implementing this novel model into existing infrastructures, one can greatly improve the performance of generated and transmitted messages. Therefore, this research confirms the feasibility of improving the data contained in the cooperative messages sent, promoting the advancement of cooperative driving technologies, and ensuring the promotion of safe connected vehicles and more effective cooperative intelligent transportation systems.