Real-Time Glass Detection and Reprojection using Sensor Fusion Onboard Aerial Robots

It verifies that the space around the detected speckle is empty. To do this efficiently, an integral image of the binarized depth map is computed, which allows for rapid, constant-time queries of the pixel sum within any rectangular region. We check the pixel sum in eight rectangular regions surrounding the speckle's bounding box. If the ratio of filled pixels to total pixels within these regions is below a low threshold (e.g., 0.07), the speckle is considered isolated within a glass plane. T emporal Consistency: A final filter operates on a tracking-by-detection principle to ensure identified features are persistent and not transient sensor noise. A speckle is confirmed and passed to the mapping algorithm only after its required count (e.g., 1-3 detections) is exceeded across multiple consecutive frames. To prevent the accumulation of false positives and old detections, a max age parameter is used to expire and remove tracks that have not been seen for a specified duration. D. Transparent Plane Reprojection The final stage of our methodology involves segmenting empty regions in the depth map and reprojecting the confirmed transparent planes. The algorithm first identifies the empty regions in the depth image and applies a non-maximum suppression (NMS) algorithm to merge redundant empty regions, ensuring a single, accurate representation of each transparent plane.