park and carla

This experiment addresses the safety challenges of intelligent devices in hazardous areas without protective measures in real park environments. Due to the lack of publicly available datasets, this study constructed a dedicated dataset to evaluate the effectiveness of the proposed method. The dataset comprises real-world images from actual scenes and simulated images extracted from a simulator. The annotations include two classes: drivable areas and background regions.

The real-world images were extracted and filtered from the Summer Palace and Beihai Park video footage, resulting in 583 selected road scene images. To prevent overfitting and enhance the generalization capability of the model, data augmentation techniques such as random flipping, scaling, brightness adjustment, and contrast enhancement were applied. The processed dataset was expanded to 1,166 images. The simulated images were generated in the Carla simulator, using scenes from Town1, Town6, and Town7, which closely resemble park environments. A total of 11,025 raw images and semantic segmentation images were obtained. Pixel-level segmentation labels were processed to ensure consistency with the annotations of the real-world images. The dataset comprises a total of 12,191 images, all uniformly resized to a resolution of 1024×512 pixels. In order to ensure the reproducibility of our results, the dataset was divided into training, validation, and testing sets in a standard ratio of 7:1.5:1.5.