VariationalVision

VariationalVision is an anomaly-detection system designed to identify unusual driving scenarios for self-driving cars. Using the CARLA simulator, the project collected 50,000 training images via an autopilot mode and performed extensive data augmentation to grow the dataset to over 100,000 samples.

At its core, the model is a convolutional variational autoencoder trained with KL divergence loss and reconstruction loss, enabling it to distinguish normal driving scenes from anomalous ones. To fine-tune the classification, a ResNet model was employed to detect unusual scenes. Training took place on an A100 GPU in the cloud, while inference ran locally on an RTX 3050 Ti.

Throughout the process, techniques like kernel density estimation (KDE) and spatial downsampling were used to reduce dimensions and capture critical features more efficiently. Written in PyTorch, the system aims to enrich the training datasets for self-driving cars by incorporating edge-case scenarios—particularly relevant in regions like South Asia where atypical road conditions and driving behaviors are prevalent. By spotting anomalies effectively, VariationalVision promotes safer navigation strategies, allowing autonomous vehicles to adapt faster to challenging, unfamiliar situations.