Construction and Validation of an Editable Image Dataset for Black Hole Effect Scenes in YOLO Object Detection

The YOLO object detection algorithm is widely applied in intelligent assisted driving systems. However, existing datasets lack targeted coverage of the black hole effect scenarios at bridge and tunnel entrances. To address this issue, this paper constructs a black hole dataset, which consists of two subsets, namely BH-Normal and BH-Blackhole, involving four typical scenarios, namely suburban beltway tunnels, urban underground tunnels, railway overpasses, and interchanges, as well as multiple influencing factors such as different illumination angles and different distances from tunnel entrances. The dataset contains 27,868 labeled images in YOLO format, with 71,956 labeled instances. Based on statistical analysis of full-day time-series and distance-sequence data, this paper reveals the evolution pattern of black hole effect intensity at tunnel entrances with respect to observation distance and time of day. Meanwhile, this paper constructs an editable black hole dataset and develops a one-stop YOLO dataset creation and management toolkit. Baseline validation using mainstream full-scale YOLO models shows that the model fine-tuned solely on BH-Normal suffers an average [email protected]:0.95 drop of over 20% under black hole effect conditions; with the introduction of BH-Blackhole and secondary fine-tuning, the model’s [email protected]:0.95 is relatively improved by more than 14.54%. This confirms that the black hole dataset can effectively enhance the target detection performance of YOLO models in black hole effect scenarios at bridge and tunnel entrances. Furthermore, the baseline results of the YOLO model obtained in the black hole dataset can provide benchmark support for optimizing intelligent assisted driving visual perception models.