BEV-MODNet: Monocular Camera based Bird's Eye View Moving Object Detection for Autonomous Driving

Detection of moving objects is a very important task in autonomous driving\nsystems. After the perception phase, motion planning is typically performed in\nBird's Eye View (BEV) space. This would require projection of objects detected\non the image plane to top view BEV plane. Such a projection is prone to errors\ndue to lack of depth information and noisy mapping in far away areas. CNNs can\nleverage the global context in the scene to project better. In this work, we\nexplore end-to-end Moving Object Detection (MOD) on the BEV map directly using\nmonocular images as input. To the best of our knowledge, such a dataset does\nnot exist and we create an extended KITTI-raw dataset consisting of 12.9k\nimages with annotations of moving object masks in BEV space for five classes.\nThe dataset is intended to be used for class agnostic motion cue based object\ndetection and classes are provided as meta-data for better tuning. We design\nand implement a two-stream RGB and optical flow fusion architecture which\noutputs motion segmentation directly in BEV space. We compare it with inverse\nperspective mapping of state-of-the-art motion segmentation predictions on the\nimage plane. We observe a significant improvement of 13% in mIoU using the\nsimple baseline implementation. This demonstrates the ability to directly learn\nmotion segmentation output in BEV space. Qualitative results of our baseline\nand the dataset annotations can be found in\nhttps://sites.google.com/view/bev-modnet.\n