evo-egomotion-dynamic-analysis

Analyzes consecutive grayscale frame pairs to classify camera motion and detect dynamic objects.

Key Functions

• compute_egomotion_labels(prev_gray, curr_gray, th_trans=2.0, th_scale=0.015, th_rot=0.01) Returns (labels, good_prev, good_curr) where labels is a list of motion strings.

• detect_dynamic_objects(prev_gray, curr_gray, good_prev, good_curr) Returns boolean mask (H,W) where True = dynamic pixel.

• analyze_frame_pairs(frames) Processes all consecutive pairs. Returns (motion_labels, dynamic_masks). motion_labels: list of (sample_idx_prev, sample_idx_curr, labels_list) dynamic_masks: list of boolean 2D arrays

Usage

import sys
sys.path.insert(0, '/app/environment/skills/evo-egomotion-dynamic-analysis/scripts')
from utils import analyze_frame_pairs

motion_labels, dynamic_masks = analyze_frame_pairs(grayscale_frames)

Motion Classification Rules

• Dolly In/Out: scale deviation > 0.015 from 1.0
• Roll Left/Right: rotation > 0.01 rad
• Pan Left/Right: |tx| > 2.0 px AND |tx| >= |ty|. Positive tx = Pan Left.
• Tilt Up/Down: |ty| > 2.0 px AND |ty| > |tx|. Positive ty = Tilt Up.
• Stay: if no other label triggered

Pipeline

1. Shi-Tomasi corners (goodFeaturesToTrack, maxCorners=300)
2. Lucas-Kanade optical flow (calcOpticalFlowPyrLK)
3. estimateAffinePartial2D with RANSAC for egomotion
4. findHomography with RANSAC for motion compensation
5. warpPerspective + absdiff + threshold(25) for dynamic mask
6. Morphological open(3x3) + close(7x7) cleanup