This dataset was collected for my senior year project titled "Developing an AI-Powered Autonomous Delivery Vehicle For Efficient Last Mile Delivery" at the Lahore University of Management Sciences (LUMS). The objective of this project was to develop an ADV that can self-navigate its way across the LUMS campus roads. More can be read about this project in this report.
Visual perception plays a critical role in the functionality of Autonomous Delivery Vehicles (ADVs), as they must accurately identify and differentiate between various objects in their environment. Achieving this requires real-time semantic segmentation, which we achieved by deploying our trained model on an NVIDIA Jetson Nano. The effectiveness of the segmentation model is dependent on its training with a robust and comprehensive dataset. To meet this requirement, an extensive was meticulously gathered and manually annotated using the RoboFlow platform.
The dataset consists of 5 classes:
- 0 - background
- 1 - obstacle
- 2 - road
- 3 - sidewalk
- 4 - speedbreaker
The original dataset consisted of 926 unique images. However, the dataset was increased to 2224 images after we applied the following augmentations to create 3 versions of each source image for increasing robustness:
- 50% probability of horizontal flip
- Randomly crop between 0 and 30 percent of the image
- Random rotation of between -15 and +15 degrees
- Random brigthness adjustment of between -25 and +25 percent
- Random exposure adjustment of between -15 and +15 percent
- Random Gaussian blur of between 0 and 2 pixels
- Salt and pepper noise was applied to 1.05 percent of pixels
Here are 2 sample images from the dataset with their corresponding masks:
