Abstract:
The pedestrian safety of urban traffic traffic is an issue of serious concern, particularly in the mixed environment where there is a combination of man operated and automated vehicles on the road. This thesis aims at constructing and testing a camera-based pedestrian detection unit, which will be able to assist automated vehicles to take emergency-braking decisions. The model used is a YOLOv8n, which is trained on the Joint Attention in Autonomous Driving (JAAD) dataset, in which video clips are transformed into single-class pedestrian detection labeled frames. The training pipeline is trained on a local workstation with an RTX 3050 and based on the standard object-detection metrics, such as mAP@50, mAP@50–95, precision, recall, F1-score and confusion-matrix measures, to measure results.A trained detector on JAAD has a mAP 50 of 94.24, mAP 5095 of 70.68, precision of 95.97, recall of 87.21, and an F1-score of 91.38, indicating that a small YOLOv8n model can be useful in the task of pedestrian detection in real-life driving scenarios. To illustrate that this perception module can be applied to a simulated driving stack, the model is linked to the CARLA simulator and the detections in various urban settings, including the crosswalks, parking lots and the streets, are visualized as bounding boxes. Despite the fact that the current system does not adopt a complete braking controller, the findings indicate that the proposed detector is a good basis of future vision-based emergency-braking pipelines and more complex decision making in autonomous vehicles.
