Autonomous vehicles use a variety of software-based detection algorithms to detect the static and dynamic objects in their surroundings. Detected objects are then used as bases for the higher-level tasks required for the safe operation of the vehicles. Due to their abundance in most road environments, poles are some of the key static landmarks used for the tasks of autonomous vehicle mapping and localization. However, most of the work done in the field of pole-like object detection has been focused on the use of LiDAR-generated 3D point clouds. This study explores the validity of monocular camera-based pole detection approaches as means of redundancy for environmental perception. It uses the YOLOv4 object detector combined with the MiDaS depth estimator to ultimately produce pole detections on inverse depth maps. To train and test the YOLOv4 model, a new 3000-image dataset based on BDD100K was created by manually generating accurate pole ground truth labels. The final model produced by the study is capable of predicting poles with an mAP of over 71%.
Detection of Vertical Poles in a Road Environment Using Monocular Images based on YOLOv4
2022-11-14
718867 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Pavement Crack Detection from UAV Images Using YOLOv4
Springer Verlag | 2023
|Pedestrian Detection Algorithm Based on Improved YOLOv4
IEEE | 2023
|Accurate Real-time Ship Target detection Using Yolov4
IEEE | 2021
|