Object Tracking MCQ
Video tracking basics, Kalman filtering, SORT/DeepSORT, and modern multi-object tracking ideas.
Object Tracking Basics MCQ
Object tracking
Tracking maintains identities of objects over time—single-object (SOT) focuses on one target; multi-object (MOT) handles many with births, deaths, and occlusions.
Association
Match current detections to existing tracks using motion prediction, appearance, and optimization (Hungarian, network flow).
Challenges
Motion
Constant velocity, Kalman prediction, or learned motion for gating candidates.
Appearance
Re-ID embeddings help after occlusion—DeepSORT style.
MOTA / IDF1
Classic MOT metrics mix detection accuracy and identity consistency.
Online
Causal processing for robotics; offline can use future frames globally.
Loop
Predict state → associate detections → update tracks
Kalman Filter MCQ
Kalman filter
For linear dynamics and Gaussian noise, the Kalman filter is the optimal recursive estimator. In CV, constant-velocity models smooth bounding boxes or keypoints between detections.
Two steps
Predict: propagate mean and covariance with F, Q. Update: fuse measurement with H, R via Kalman gain.
Details
State vector
Often position+velocity per axis; higher order adds acceleration (constant acceleration model).
Q vs R
Process noise Q allows model mismatch; measurement noise R trusts observations.
Nonlinear
EKF/UKF/particle filters extend when motion or measurement models are nonlinear.
In tracking
Prediction gates association; update pulls estimate to matched detection.
Recursion
x̂_{k|k-1} → z_k → x̂_{k|k}
SORT & DeepSORT MCQ
SORT and DeepSORT
SORT demonstrated that a simple Kalman filter + Hungarian assignment on IoU costs achieves surprising MOT speed/accuracy when paired with a strong detector. DeepSORT adds a cosine-distance appearance metric to re-identify tracks after occlusion.
Cascade matching
DeepSORT matches recently seen tracks to high-confidence detections first using appearance, then IoU for remaining—reduces fragmentation.
Practical notes
Cost matrix
1 − IoU or Mahalanobis gate + embedding distance—threshold unassigned pairs.
Age / hits
Track management rules delete stale tracks and confirm new ones after hits.
Gallery
DeepSORT keeps short appearance history per track for robust matching.
Speed
SORT is very fast; embedding extraction adds compute but still real-time on GPU.
Pipeline
Detect → Predict KF → Build costs → Hungarian → Update confirmed tracks