Weights for Autonomous Campus Navigation
Description
This release provides high-performance pre-trained model weights to support autonomous path tracking research using deep learning. The weights correspond to object detection models trained on the AutoNaVIT dataset and are ideal for benchmarking, transfer learning, and validation tasks in autonomous navigation systems. Key Components: Pre-trained Model Weights Only – No image or annotation files are included in this release Models Provided: YOLOv8n (Nano version) YOLOv8s (Small version) YOLOv5s (Small version) Performance Metrics: All models demonstrated mean Average Precision (mAP) > 95% when evaluated on the full AutoNaVIT dataset Annotation Format Used in Training: CSV format capturing object classification and bounding box coordinates for the following classes: Kerb Obstacle Path These models were trained to recognize key visual cues necessary for autonomous path tracking, such as kerbs, obstacles, and navigable paths. The provided weights are optimized for high accuracy and fast inference, and they can be integrated into diverse object detection pipelines. Note: This release includes only the model weights. The annotated images and label files used for training are part of a separate dataset release.
Files
Steps to reproduce
1) Collect Images Capture video or take images using a high-resolution camera (e.g., Sony IMX890) along urban or semi-urban roads. Make sure the images include common elements like kerbs, paths, and obstacles. Daylight conditions are preferred for clarity. 2) Extract Frames (Optional) If you're using a video, extract frames at around 2 FPS to get a good spread of varied scenes. Maintain the original resolution for best detection results. 3) Annotate for Inference (Optional) If needed, annotate your images using tools like Roboflow or LabelImg. Save the annotations in CSV format with object labels and bounding box coordinates. This step is optional if you’re only using the provided weights and annotations. 4) Use Pre-Trained Weights Download and load the pre-trained weights provided for: YOLOv8n YOLOv8s YOLOv5s These models are already trained and achieve over 95% mAP, making them ideal for direct inference or transfer learning. 5) Run Inference or Fine-Tune You can either: Use the models directly for object detection on new images Fine-tune the models on your own annotated dataset to adapt to specific conditions or environments 6) Integrate into Workflow Use popular frameworks like PyTorch (Ultralytics YOLO) or TensorFlow to run inference, visualize results, or integrate the models into a larger autonomous navigation system.
Institutions
- VIT University - Chennai Campus