Unsupervised Motion Segmentation for Neuromorphic Aerial Surveillance

Frame-based vision sensors on aerial platforms face bandwidth-latency challenges. On the contrary, event cameras offer promise with their high temporal resolution and low power needs. Existing methods for event-based motion segmentation often require manual labeling or scene-specific parameter tuning. Our approach addresses these issues by employing self-supervised transformers on event data and optical flow, eliminating the need for human annotations and reducing parameter tuning. We evaluate our method extensively on diverse datasets using an HD event camera on a dynamic aerial platform in urban settings, demonstrating superior performance compared to existing methods in handling various motion types and multiple moving objects