Overview
Syllabus
Intro
Tracking: problem & data Online tracking: following the location of one target in video starting from a bounding box in the first frame by computationally exploiting the temporal dimension .
Amsterdam Library Of Videos ALOV contains more than 300 videos, mostly with one event per video, annotated the ground truth/Sth frame.
NCC Normalized Cross-Correlation Direct target matching by normalized cross-correlation Intensity values in the initial target box as template. Sampling uniformly around the previous position
LKT Lucas Kanade Tracker Affine-transformed match between the target and candidates. Around the previous location by incremental image alignment. Spatiotemporal derivatives Warping around the previous position
MST Mean Shift Tracker Matching with RGB-histogramDy the Bhattacharyya metric The location by mean shift maximizes the Bhattacharyya modus. Target histogram formed in the first frame No update
IVT Incremental Visual Tracking Extended appearance model captures the past i Figen Images Stored in a leaking memory. Similarity by subspace distance. Gaussian particle filtering sampling around the previous The candidate window with the minimum score is selected
TST Tracking by Sampling Trackers Sampling many IT-like trackers as an extended model The state stores the center, scale and spatial Multiple locations and scales, filters, motion in multiple Gaussians. The best target state is selected from the space of trackers.
The tracker trains a linear discriminant classifier as only the difference between the object and the background counts. SURF texture features from target against the local background. Updated by a leaking memory on the training data.
Updating the FBT-model The foreground template is updated in every frame as the weighted average between the old template and the object feature in the current frame
HBT Hough-Based Tracking Discriminative classifier on Lab-color, gradients and positions The Hough Forest provides a probability map of the target area. Back projection aims to avoid inclusion of the wrong labels Segment the target using grabcut and hence generate new samples.
TLD Tracking, Learning and Detection 50 Top detections on LBPs and KLT optical flow are combined by NCC. Samples are selected in, around and away from the target to update. If neither of the two trackers outputs, TD declares loss and recovers.
Shadows The effect of shadows. Heavy shadow has an impact almost for all.
Systematic overview of trackers Trackers have to find an integral solution on the basis of 1. a spotial representation, 2. an appearance representation, 3. a motion model 4. an inference method to get to the next state, 5. a method for updating the internal models.
Feature types Essentially different types of features: 1. Geometric 2D modeloverall locality matters 2. Histogram 10 array-spatially disturbed, sufficient identification 3. Feature OD vector = feature values provide sufficient match Interesting extension for known shapes 0. 3D Geometric model Rita's group 2012
Updating revisited 1. No updating does remarkably well do not over-update. 2. Update at least for scale, rotation and occlusion 3. Focusing on one condition ruins another broad dataset
The hardness of tracking Tracking aims to learn a target from the first few pictures; the target and the background may be dynamic in appearance, with unpredicted motion, and in difficult scenes.
Taught by
UCF CRCV