Edge-Embedded Ensemble Model for High Accuracy and Real-Time Lane Detection

In this study, we propose an ensemble model for accurate lane detection (LD).The proposed model combines the outputs of two branches, leveraging the strengths of Convolutional Neural Networks (CNN), spatial CNN (SCNN), and Swin Transformers (ST).The CNN components focus on extracting local features to identify finer details of the lane, whereas the SCNN improves the model's capacity to capture spatial dependencies among multiple pixels in various areas of the image.The attention mechanisms of the ST capture global context and long-range dependencies.Consequently, the ensemble model significantly enhances the overall LD performance.Experiments were conducted using Tusimple and BDD100K datasets.The model achieved an F1 score of 98.23%, accuracy of 98.8%, and IoU of 36.5%.The results of its integration on the NVIDIA Jetson Xavier demonstrate its suitability for edge-based real-time applications.The source code is accessible at: https://github.com/mustapha930/Lane_detection.