Tracking and integrated navigation Konrad Schindler Institute of Geodesy and Photogrammetry
Tracking Navigation needs predictions for dynamic objects estimate trajectories in 3D world coordinates and extrapolate [Ess, Leibe, Schindler, Van Gool; ICRA 10] 2
Tracking Ego-motion in world coordinates known from navigation photogrammetry ( visual odometry ) or other sensors 3
Stereo: place detected objects in 3D context Ground plane Depth map 4
Stereo: objects in 3D context groundplane object detection o i o i i =1..n d i image I range image D object depth d i I D O range confidence O groundplane observation object detection verification by depth 5
Stereo: objects in 3D context Result: Objects in 3D space for a single time step 6
Multi-object Tracking Connecting the dots through time: tracking combine detector evidence with assumptions about target dynamics Single-target tracking: recursive methods Kalman filter Particle filter (sequential Monte Carlo) 7
Multi-object Tracking Data association problem how to make sure we link the right objects through time? Possible solution: hypothesize-and-verify link all physically plausible candidate trajectories select best non-redundant subset by discrete optimization 3D 8
Multi-object Tracking trajectory model motion: extended Kalman filter radiometry: colour histogram Candidate trajectories Selected trajectories 9
Multi-object Tracking How does tracking help? many false positives (errors of commission) suppressed most false negatives (errors of omission) filled in objects have ID and history prediction 10
Putting it all together Detection 3D localization tracking prediction 11
Putting it all together Detection 3D localization tracking prediction 12
Putting it all together Detection 3D localization tracking prediction 13
Putting it all together Detection 3D localization tracking prediction 14
Tracking: feedback to geometry Photogrammetry assumes a static scene moving objects (potentially) violate that assumption mask out corresponding regions more reliable orientation feedback detector feedback Tracker Predict position Detect features 3D-2D matching Pose estimation Update position Triangulate Bundle adjust 15
Tracking: feedback to geometry uncertainty depends on tiepoint distribution again assuming a static scene no masking: variance underestimated with masking: correct error ellipsoid 16
Tracking: feedback to geometry naive orientation with error diagnosis diagnosis and masking 17
Results (cooperation with Toyota Motor Corp.) 18
Results: Quantitative Evaluation Length [s] # Persons > 80% Tracked < 20% Tracked False Alarm/frm # ID Switches Latency [frames] 75 89 72% 15% 0.62 16 1.5 60 107 64% 25% 1.09 6 2 A notorious problem: evaluating tracking high latency in the far field what is the gold standard? suitable quality measures for different applications? how to acquire reference data for pedestrian locations? 19
Results (cooperation with ETHZ robotics group) 20
Results: Typical Mistakes Model errors of object detector too big people too small people partial occlusions False positives in trees, street furniture,... (for navigation purposes a graceful degradation) 21
Outlook - autonomous driving Will robots soon drive well enough for public roads? Video: Stanford Automotive Innovation Laboratory 22