Estimating the Driving State of Oncoming Vehicles From a Moving Platform Using Stereo Vision IEEE Intelligent Transportation Systems 2009 M.S. Student, Heejong Hong Alexander Barth and Uwe Franke
Introduction Related Works Proposed Method Experimental Results Conclusion Outline 2
Introduction Driver-assistance and safety systems Dynamic Object Detection for DAS Safety System with Dynamic Path Estimation
1. A model-free object representation based on groups 2. Fusion active sensors 3. Track-before-detection 4. Rediscovering an image region labeled as vehicle Related Works 4 D. Beymer, P. McLauchlan, B. Coifman and J. Malik "A real-time computer vision system for measuring traffic parameters", Proc. Comput. Vis. Pattern Recog, pp M. Maehlisch, W. Ritter and K. Dietmayer "De-cluttering with integrated probabilistic data association for multisensor multitarget ACC vehicle tracking", Proc. IEEE Intell. Veh. Symp., pp U. Franke, C. Rabe, H. Badino and S. Gehrig "6D-vision: Fusion of stereo and motion for robust environment perception", Proc. 27th DAGM Symp., pp X. Li, X. Yao, Y. Murphey, R. Karlsen and G. Gerhart "A real-time vehicle detection and tracking system in outdoor traffic scenes", Proc. 17th Int. Conf. Pattern Recog., pp.II:761 -II:
5 Proposed Method
Object Model 1. Pose (relative orientation and translation to ego-vehicle) 2. Motion State (velocity, acceleration, yaw rate) 3. Shape (rigid 3-D point cloud) Motion State Shape Pose
Object Tracking Extended Kalman Filter (EKF) : Kalman filter for nonlinear model State transition(f) and observation model(h) Discrete-time predict and update equations Wikipedia : Jacobian of system & measurement model Example)
Object Tracking 1. State Vector of an object instance Reference point in ego-coordinates Rotation point in object-coordinates The object origin is ideally defined on the center rear axis
Object Tracking 2. Dynamic(System) Model R is 3x3 rotation matrix around the height axis N. Kaempchen, K. Weiss, M. Schaefer and K. Dietmayer "IMM object tracking for high dynamic driving maneuvers", Proc. IEEE Intell. Veh. Symp., pp Predicted state vector Time-discrete system EquationTransformation of an object point Translation matrix
Object Tracking 3. Measurement Model The measurement nonlinear eq. : perspective camera model Jacobian of measurement model Objects feature points on image coordinates using feature tracker (KLT) Feature point tracking using KLT
Kalman Filter Initialization 1. Image Based Initialization 2. Radar-Based Initialization (detect oncoming vehicle up to 200m) The mean velocity vector : The centroid of the 3-D positions : Initial Yaw : The lateral and longitudinal positions of the radar target :, Absolute radar velocity of the object : Initial Yaw :
Point Model Update 1. Maximum-likelihood estimation 2. Simple average filter Object’s Shape Expectation = 3x3 covariance matrix of t Expected object’s shape
13 Experimental Result
Simulation Results Synthetic Sequence
Real World Results Country Road Curve I
Real World Results Country Road Curve II
Real World Results Oncoming Traffic at Intersections
Real World Results Leading Vehicles & Partial Occlusions
Real World Results Challenges and Limits
Joint Histogram Based Cost Aggregation for Stereo Matching - TPAMI Conclusion
Joint Histogram Based Cost Aggregation for Stereo Matching - TPAMI Contribution New method for the image-based real-time tracking (25Hz, 640x480) Results of experiments with synthetic data & real-world Two different object detection method (image & radar) Feature-based object point model does not require a priori knowledge about the object’s shape Weakness No specific system block diagram User defined rotation point Shape depends on outlier removing algorithm (ex : max distance parameter) Shape is very sensitive about outlier of point cloud (because of yaw)
22 Thank you!