BEV-MoSeg: Segmenting Moving Objects in Bird's Eye View

Accurate detection of moving objects plays a vital role in motion planning and vehicle maneuvering for autonomous vehicles. Though there is a significant improvement in perception tasks like object detection and semantic segmentation by adopting Bird's Eye View (BEV) based techniques like LiftSplatShoot, SimpleBEV etc., the moving object segmentation has gained limited attention. This research addresses this gap and propose a novel end-to-end architecture that implicitly utilizes temporal cues like optical flow in BEV space by correlation or cross-attention for moving vehicle segmentation. This work also introduces custom labels to annotate moving objects in the nuScenes dataset, enhancing its utility for the BEV motion segmentation task. We achieved an Moving Vehicle IoU Score of 26% on nuScenes dataset on full six camera rig and 22% on single front camera. The code for generating these labels and the qualitative results of our model can be found in, Project page with code: https://ajayrafa25.github.io/BEV- MoSeg/