1. Object Tracking and Detection after Occlusion via Numerical Hybrid Local and Global Mode-seeking
Zhaozheng Yin and Robert T. Collins Dept. of Computer Science and Engineering, The Pennsylvania State University, USA
Introduction
Detection by Global Mode-seeking
Experimental Results
We seek to harness the power of generic optimization libraries to perform appearance-based tracking via numerical mode-seeking. We consider two cases: local mode-seeking for frame-to-frame tracking, and global mode-seeking for object detection after occlusion or tracking failure.
To recover from occlusion, we consider object detection as a global optimization problem and use Adaptive Simulated Annealing (ASA). As a Monte Carlo approach, ASA stochastically samples the parameter space, allowing it to avoid being trapped at local modes. We apply cluster analysis on the sampled parameter space to redetect the object and renew the local tracker.
Local Mode-seeking under Normal Tracking Conditions:
Figure 9. Top row: the face is represented by HoG and tracked by the Simplex method using EMD distance. Bottom row: the hand is represented by color histogram and tracked by the Simplex method using Bhattacharya coefficient.
Figure 5. Detecting an object in 2D translation space for a given (theta; s). The top row shows the detection results with white boxes representing the modes found by ASA. The bottom row shows corresponding sampling maps where the bright pixels represent the sampled points with their objective function values. From left to right: ASA detection restarted from 25, 16, 4 and 1 different initial points, respectively.
Tracking by Local Mode-seeking
Numerical Hybrid Local and Global Mode-seeking
Our optimization function f is a black-box function that encodes an object appearance model and a model-to-image similarity measure. Input values to f are the model pose parameters, e.g. location (u,v), orientation theta and scale s, and the output is a scalar value measuring model-to-image similarity for that pose. The object is tracked to the next frame by directly seeking the argmax of objective function f(u, v, theta, s) in 4D parameter space.
ASA detection in 2D translation space:
ASA detection in 3D space (translation plus scale):
Figure 11. Left two columns: airborne car chase videos; Right two columns: VIVID benchmark datasets EgTest05 and PkTest01.
ASA detection in 4D space (translation, rotation and scale):
Figure 4. Dashed polygons represent different starting positions. Solid polygons are the modes found by local mode-seeking methods. The solid lines show the search paths. (a) The reference appearance model is represented by the color histogram inside the polygon; (b) Steepest ascent method; (c) Trust region algorithm; (d) Nelder-Mead simplex algorithm.