From 34844fe6eafd13874ccdd05030fca595176403f7 Mon Sep 17 00:00:00 2001 From: Théo de la Hogue Date: Mon, 19 Sep 2022 21:43:33 +0200 Subject: Refactoring warnings. Separating video frame processing and visualisations. Improving aruco tracking into focus area only. Making focus area rectangular. Ignoring frame where hed is moving. --- .../export_tobii_segment_aruco_visual_scan.py | 220 +++++++++++++-------- 1 file changed, 133 insertions(+), 87 deletions(-) diff --git a/src/argaze/utils/export_tobii_segment_aruco_visual_scan.py b/src/argaze/utils/export_tobii_segment_aruco_visual_scan.py index 4825c9e..653a5fa 100644 --- a/src/argaze/utils/export_tobii_segment_aruco_visual_scan.py +++ b/src/argaze/utils/export_tobii_segment_aruco_visual_scan.py @@ -95,6 +95,9 @@ def main(): # Access to timestamped gaze 3D positions data buffer tobii_ts_gaze_positions_3d = tobii_segment_data['GazePosition3D'] + # Access to timestamped head rotations data buffer + tobii_ts_head_rotations = tobii_segment_data['Gyroscope'] + # Prepare video exportation at the same format than segment video output_video = TobiiVideo.TobiiVideoOutput(vs_video_filepath, tobii_segment_video.get_stream()) @@ -199,140 +202,183 @@ def main(): # Initialise progress bar #MiscFeatures.printProgressBar(0, tobii_segment_video.get_duration()/1000, prefix = 'Progress:', suffix = 'Complete', length = 100) + head_moving = False + head_movement_last = 0. + # Iterate on video frames for video_ts, video_frame in tobii_segment_video.frames(): video_ts_ms = video_ts / 1000 - - # Track markers with pose estimation and draw them - aruco_tracker.track(video_frame.matrix) - aruco_tracker.draw(video_frame.matrix) + visu_frame = video_frame.copy() # Write segment timing - cv.putText(video_frame.matrix, f'Segment time: {int(video_ts_ms)} ms', (20, 40), cv.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 1, cv.LINE_AA) - + cv.putText(visu_frame.matrix, f'Segment time: {int(video_ts_ms)} ms', (20, 40), cv.FONT_HERSHEY_SIMPLEX, 1, (127, 127, 127), 1, cv.LINE_AA) + + # Draw focus area + cv.rectangle(visu_frame.matrix, (int(video_frame.width/6), 0), (int(video_frame.width*(1-1/6)), int(video_frame.height)), (255, 150, 150), 1) + + # Draw center + cv.line(visu_frame.matrix, (int(visu_frame.width/2) - 50, int(visu_frame.height/2)), (int(visu_frame.width/2) + 50, int(visu_frame.height/2)), (255, 150, 150), 1) + cv.line(visu_frame.matrix, (int(visu_frame.width/2), int(visu_frame.height/2) - 50), (int(visu_frame.width/2), int(visu_frame.height/2) + 50), (255, 150, 150), 1) + + # Process video and data frame try: + # Get nearest head rotation before video timestamp and remove all head rotations before + _, nearest_head_rotation = tobii_ts_head_rotations.pop_first_until(video_ts) + + # Calculate head movement considering only head yaw and pitch + head_movement = numpy.array(nearest_head_rotation.value) + head_movement_px = head_movement.astype(int) + head_movement_norm = numpy.linalg.norm(head_movement[0:2]) + + # Draw movement vector + cv.line(visu_frame.matrix, (int(visu_frame.width/2), int(visu_frame.height/2)), (int(visu_frame.width/2) + head_movement_px[1], int(visu_frame.height/2) - head_movement_px[0]), (150, 150, 150), 3) + + # Head movement detection hysteresis + # TODO : pass the threshold value as argument + if not head_moving and head_movement_norm > 50: + head_moving = True + + if head_moving and head_movement_norm < 10: + head_moving = False + + # Ignore frame where head is moving + if head_moving: + raise UserWarning('Head is moving') + # Get nearest gaze position before video timestamp and remove all gaze positions before _, nearest_gaze_position = tobii_ts_gaze_positions.pop_first_until(video_ts) + gaze_position_pixel = (int(nearest_gaze_position.value[0] * visu_frame.width), int(nearest_gaze_position.value[1] * visu_frame.height)) + + # Draw gaze position + cv.circle(visu_frame.matrix, gaze_position_pixel, 2, (0, 255, 255), -1) + # Get nearest gaze position 3D before video timestamp and remove all gaze positions before _, nearest_gaze_position_3d = tobii_ts_gaze_positions_3d.pop_first_until(video_ts) - # Consider gaze position if gaze precision can be evaluated - if nearest_gaze_position_3d.value[2] > 0: + # Ignore frame when gaze precison can't be evaluated + if nearest_gaze_position_3d.value[2] <= 0: + raise UserWarning('Negative Z gaze position 3D value') - gaze_position_pixel = (int(nearest_gaze_position.value[0] * video_frame.width), int(nearest_gaze_position.value[1] * video_frame.height)) + gaze_accuracy_mm = numpy.tan(numpy.deg2rad(tobii_accuracy)) * nearest_gaze_position_3d.value[2] + tobii_camera_hfov_mm = numpy.tan(numpy.deg2rad(tobii_camera_hfov / 2)) * nearest_gaze_position_3d.value[2] + gaze_accuracy_pixel = round(visu_frame.width * float(gaze_accuracy_mm) / float(tobii_camera_hfov_mm)) - gaze_accuracy_mm = numpy.tan(numpy.deg2rad(tobii_accuracy)) * nearest_gaze_position_3d.value[2] - tobii_camera_hfov_mm = numpy.tan(numpy.deg2rad(tobii_camera_hfov / 2)) * nearest_gaze_position_3d.value[2] - gaze_accuracy_pixel = round(video_frame.width * float(gaze_accuracy_mm) / float(tobii_camera_hfov_mm)) + # Draw gaze accuracy + cv.circle(visu_frame.matrix, gaze_position_pixel, gaze_accuracy_pixel, (0, 255, 255), 1) - # Draw gaze position and accuracy - cv.circle(video_frame.matrix, gaze_position_pixel, 2, (0, 255, 255), -1) - cv.circle(video_frame.matrix, gaze_position_pixel, gaze_accuracy_pixel, (0, 255, 255), 1) + # Store gaze position and precision at this time in millisecond + ts_gaze_positions[round(video_ts_ms)] = gaze_position_pixel + ts_gaze_accuracies[round(video_ts_ms)] = gaze_accuracy_pixel - # Store gaze position and precision at this time in millisecond - ts_gaze_positions[round(video_ts_ms)] = gaze_position_pixel - ts_gaze_accuracies[round(video_ts_ms)] = gaze_accuracy_pixel + # Hide frame left and right borders before tracking to ignore markers outside focus area + cv.rectangle(video_frame.matrix, (0, 0), (int(video_frame.width/6), int(video_frame.height)), (0, 0, 0), -1) + cv.rectangle(video_frame.matrix, (int(video_frame.width*(1 - 1/6)), 0), (int(video_frame.width), int(video_frame.height)), (0, 0, 0), -1) - else: + # Track markers with pose estimation and draw them + aruco_tracker.track(video_frame.matrix) + aruco_tracker.draw(visu_frame.matrix) - ValueError('Unable to evaluate gaze precision') + # Project 3D scene on each video frame and the visualisation frame + if aruco_tracker.get_markers_number(): - # Wait for gaze position - except ValueError: - continue + # Store aoi 2D video for further scene merging + aoi2D_dict = {} - # Draw focus area - cv.circle(video_frame.matrix, (int(video_frame.width/2), int(video_frame.height/2)), int(video_frame.width/3), (255, 150, 150), 1) - - # Draw focus area center - cv.line(video_frame.matrix, (int(video_frame.width/2) - 50, int(video_frame.height/2)), (int(video_frame.width/2) + 50, int(video_frame.height/2)), (255, 150, 150), 1) - cv.line(video_frame.matrix, (int(video_frame.width/2), int(video_frame.height/2) - 50), (int(video_frame.width/2), int(video_frame.height/2) + 50), (255, 150, 150), 1) - - # Project 3D scene on each video frame and the visualisation frame - if aruco_tracker.get_markers_number(): + for (i, marker_id) in enumerate(aruco_tracker.get_markers_ids()): - # Store aoi 2D video for further scene merging - aoi2D_dict = {} + # Process marker pose + try: - for (i, marker_id) in enumerate(aruco_tracker.get_markers_ids()): + # Copy 3D scene related to detected marker + aoi3D_scene = aoi3D_scene_selector(marker_id) + + if aoi3D_scene == None: + raise UserWarning('No AOI 3D scene') - # Copy 3D scene related to detected marker - aoi3D_scene = aoi3D_scene_selector(marker_id) - - if aoi3D_scene == None: - continue - - # Ignore marker out of focus area - marker_x, marker_y = aruco_tracker.get_marker_center(i) - distance_to_center = ( (video_frame.width/2 - marker_x)**2 + (video_frame.height/2 - marker_y)**2 )**0.5 + # Transform scene into camera referential + aoi3D_camera = aoi3D_scene.transform(aruco_tracker.get_marker_translation(i), aruco_tracker.get_marker_rotation(i)) - if distance_to_center > int(video_frame.width/3): - continue + # Get aoi inside vision cone field + cone_vision_height_cm = nearest_gaze_position_3d.value[2]/10 # cm + cone_vision_radius_cm = numpy.tan(numpy.deg2rad(tobii_visual_hfov / 2)) * cone_vision_height_cm - # Transform scene into camera referential - aoi3D_camera = aoi3D_scene.transform(aruco_tracker.get_marker_translation(i), aruco_tracker.get_marker_rotation(i)) + aoi3D_inside, aoi3D_outside = aoi3D_camera.vision_cone(cone_vision_radius_cm, cone_vision_height_cm) - # Get aoi inside vision cone field - cone_vision_height_cm = nearest_gaze_position_3d.value[2]/10 # cm - cone_vision_radius_cm = numpy.tan(numpy.deg2rad(tobii_visual_hfov / 2)) * cone_vision_height_cm + # Keep only aoi inside vision cone field + aoi3D_scene = aoi3D_scene.copy(exclude=aoi3D_outside.keys()) - aoi3D_inside, aoi3D_outside = aoi3D_camera.vision_cone(cone_vision_radius_cm, cone_vision_height_cm) + # DON'T APPLY CAMERA DISTORSION : it projects points which are far from the frame into it + # This hack isn't realistic but as the gaze will mainly focus on centered AOI, where the distorsion is low, it is acceptable. + aoi2D_video_scene = aoi3D_scene.project(aruco_tracker.get_marker_translation(i), aruco_tracker.get_marker_rotation(i), aruco_camera.get_K()) - # Keep only aoi inside vision cone field - aoi3D_scene = aoi3D_scene.copy(exclude=aoi3D_outside.keys()) + # Store each 2D aoi for further scene merging + for name, aoi in aoi2D_video_scene.items(): - # DON'T APPLY CAMERA DISTORSION : it projects points which are far from the frame into it - # This hack isn't realistic but as the gaze will mainly focus on centered AOI, where the distorsion is low, it is acceptable. - aoi2D_video_scene = aoi3D_scene.project(aruco_tracker.get_marker_translation(i), aruco_tracker.get_marker_rotation(i), aruco_camera.get_K()) + if name not in aoi2D_dict.keys(): + aoi2D_dict[name] = [] - # Store each 2D aoi for further scene merging - for name, aoi in aoi2D_video_scene.items(): + aoi2D_dict[name].append(aoi.clockwise()) - if name not in aoi2D_dict.keys(): - aoi2D_dict[name] = [] + # Select 2D visu scene if there is one for the detected marker + aoi2D_visu_scene = aoi2D_visu_scene_selector(marker_id) + aoi2D_visu_frame = aoi2D_visu_frame_selector(marker_id) + + if aoi2D_visu_scene == None: + continue + + look_at = aoi2D_video_scene['Visualisation_Plan'].look_at(gaze_position_pixel) - aoi2D_dict[name].append(aoi.clockwise()) + visu_gaze_pixel = aoi2D_visu_scene['Visualisation_Plan'].looked_pixel(look_at) + cv.circle(aoi2D_visu_frame, visu_gaze_pixel, 4, (0, 0, 255), -1) - # Select 2D visu scene if there is one for the detected marker - aoi2D_visu_scene = aoi2D_visu_scene_selector(marker_id) - aoi2D_visu_frame = aoi2D_visu_frame_selector(marker_id) - - if aoi2D_visu_scene == None: - continue - - look_at = aoi2D_video_scene['Visualisation_Plan'].look_at(gaze_position_pixel) + # Write warning related to marker pose processing + except UserWarning as e: + + cv.putText(visu_frame.matrix, f'Marker {marker_id}: {e}', (20, int(visu_frame.height) - (marker_id+1) * 40), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 1, cv.LINE_AA) - visu_gaze_pixel = aoi2D_visu_scene['Visualisation_Plan'].looked_pixel(look_at) - cv.circle(aoi2D_visu_frame, visu_gaze_pixel, 4, (0, 0, 255), -1) + # Merge all 2D aoi into a single 2D scene + aoi2D_merged_scene = AOI2DScene.AOI2DScene() + for name, aoi_array in aoi2D_dict.items(): + aoi2D_merged_scene[name] = numpy.sum(aoi_array, axis=0) / len(aoi_array) - # Merge all 2D aoi into a single 2D scene - aoi2D_merged_scene = AOI2DScene.AOI2DScene() - for name, aoi_array in aoi2D_dict.items(): - aoi2D_merged_scene[name] = numpy.sum(aoi_array, axis=0) / len(aoi_array) + aoi2D_merged_scene.draw(visu_frame.matrix, gaze_position_pixel, gaze_accuracy_pixel, exclude=['Visualisation_Plan']) + + # Store 2D merged scene at this time in millisecond + ts_aois_scenes[round(video_ts_ms)] = aoi2D_merged_scene - aoi2D_merged_scene.draw(video_frame.matrix, gaze_position_pixel, gaze_accuracy_pixel, exclude=['Visualisation_Plan']) - - # Store 2D merged scene at this time in millisecond - ts_aois_scenes[round(video_ts_ms)] = aoi2D_merged_scene + else: + + raise UserWarning('No marker detected') + # Write warning related to video and data frame processing + except UserWarning as e: + + cv.putText(visu_frame.matrix, str(e), (20, 80), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 1, cv.LINE_AA) + + except ValueError: + pass + if args.window: # Close window using 'Esc' key if cv.waitKey(1) == 27: break - # Display video - cv.imshow(f'Segment {tobii_segment.get_id()} ArUco AOI', video_frame.matrix) + # Display video frame + cv.imshow(f'Segment {tobii_segment.get_id()}', video_frame.matrix) + + # Display visualisation + cv.imshow(f'Segment {tobii_segment.get_id()} ArUco AOI', visu_frame.matrix) # Display each visual scan frame - for marker_id, visu_frame in aoi2D_visu_frames.items(): - cv.imshow(f'Segment {tobii_segment.get_id()} visual scan for marker {marker_id}', visu_frame) + for marker_id, aoi2D_visu_frame in aoi2D_visu_frames.items(): + cv.imshow(f'Segment {tobii_segment.get_id()} visual scan for marker {marker_id}', visu_frame.matrix) # Write video - output_video.write(video_frame.matrix) + output_video.write(visu_frame.matrix) # Update Progress Bar progress = video_ts_ms - int(args.time_range[0] * 1000) @@ -367,8 +413,8 @@ def main(): print(f'Visual scan data saved into {vs_data_filepath}') # Export each visual scan picture - for marker_id, visu_frame in aoi2D_visu_frames.items(): - cv.imwrite(vs_visu_filepath % marker_id, visu_frame) + for marker_id, aoi2D_visu_frame in aoi2D_visu_frames.items(): + cv.imwrite(vs_visu_filepath % marker_id, visu_frame.matrix) print(f'Visual scan picture for marker {marker_id} saved into {vs_visu_filepath % marker_id}') # Notify when the visual scan video has been exported -- cgit v1.1