This paper proposes the application of a trained object detection model on different traffic scenarios and their validation based on human driver perception.
The application of the object detection model was applied to the driving scenarios, the decision tree was set up, and the results were visualized using a graphical user interface developed specifically for these purposes.
The object detection model that was applied to the driving scenarios was trained using the datasets provided by Waymo, called Waymo Open Dataset, more specifically the Perception Dataset, which contains high-resolution sensor data and labels. This dataset consists of multiple characteristics and elements, as:
– 2000 segments of 20 s each, data acquisition at 10 Hz, in different geographical zones and conditions
– 5 cameras (front and lateral) images
– 1 medium-range and 4 short-range Lidars
– 4 object classes where of interest for the object detection model have been only 3 objects: vehicle, bicycle, pedestrian.
TensorFlow Object Detection API open-source framework facilitated the development of the object detection models, where a single-shot detection (SSD) model was used due to the fact that the detection speed is more important than the model accuracy, as the continuation of this paper would be to apply the algorithm to real-time scenarios.
The objects detected by the model applied to the driving scenarios were processed and validated using the chosen decision tree. The first stage is whether the object has to be verified, then if it is too close, and then if it is in the direction of the vehicle. Based on these observations, the model warns the driver if the situation is safe or critical.
The analyzed driving scenarios were considered in a low-limit speed area with limited visibility, in city traffic during rush hour and on a winding mountain road.
The object detection model applied to the analyzed scenarios validated the results using the human driver’s perception of the situations.
Autonomous Driving Object Detection Platform
Proceedings in Automotive Engineering
International Congress of Automotive and Transport Engineering ; 2024 ; Brasov, Romania November 06, 2024 - November 08, 2024
CONAT 2024 International Congress of Automotive and Transport Engineering ; Kapitel : 27 ; 313-324
20.11.2024
12 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
3D Object Detection for Autonomous Driving
Springer Verlag | 2022
|Application of Object Detection in Autonomous Driving
IEEE | 2023
|Autonomous driving platform with independent driving wheels
Europäisches Patentamt | 2024