Computer Vision Interview 20 essential Q&A Updated 2026
Kalman

Kalman Filter for Tracking: 20 Essential Q&A

Linear Gaussian state estimation—predict with motion model, correct with measurements.

~11 min read 20 questions Advanced
predictupdateQ/REKF
1 What is the Kalman filter? 📊 medium
Answer: Optimal recursive estimator for linear systems with Gaussian noise—alternates prediction from dynamics and correction from noisy observations.
2 State-space form? 🔥 hard
Answer: x_{k+1} = F x_k + w_k (process noise), z_k = H x_k + v_k (measurement noise)—Kalman assumes linear F,H and Gaussian w,v.
3 Typical bbox state in SORT? 📊 medium
Answer: Often [cx, cy, s, r, vx, vy, vs] (center, scale area-ish, aspect, velocities)—measurements update subset.
4 Predict step? 📊 medium
Answer: x̂− = F x̂, P− = F P Fᵀ + Q—propagate mean and covariance forward in time without new measurement.
5 Update step? 📊 medium
Answer: Fuse measurement z using Kalman gain K: x̂ = x̂− + K(z − H x̂−), P = (I − K H) P−—reduce uncertainty along observed dimensions.
6 Kalman gain meaning? 🔥 hard
Answer: K balances trust in prediction vs measurement based on covariances—if R small (accurate sensor), K larger, trust measurement more.
7 Tune Q? 📊 medium
Answer: Process noise covariance—higher Q = more model uncertainty, tracker follows measurements faster but noisier.
8 Tune R? 📊 medium
Answer: Measurement noise—higher R = smoother track, lag on maneuvers; lower R = jittery if detector noisy.
9 Constant velocity model? ⚡ easy
Answer: Assumes derivative of position constant between frames—simple, works for smooth motion; fails on sharp turns.
10 Constant acceleration? 📊 medium
Answer: Adds acceleration state for more expressive motion—better for maneuvering targets, more parameters to tune.
11 When EKF? 🔥 hard
Answer: Nonlinear dynamics or measurement—linearize with Jacobians around current estimate; no longer globally optimal but widely used.
12 UKF / particle? 🔥 hard
Answer: Handle stronger nonlinearities—UKF uses sigma points; particle filters for non-Gaussian multimodal posteriors (rare in simple MOT).
13 Missing detection? ⚡ easy
Answer: Skip update; covariance grows with prediction-only steps until next match—standard in SORT when object temporarily not detected.
14 Multi-dimensional measurements? 📊 medium
Answer: H maps state to observed variables (e.g. only position observed, not velocity directly inferred from motion over time).
15 What is P? 📊 medium
Answer: State estimate covariance—uncertainty ellipsoid; should shrink after informative updates.
16 Initialize velocity? ⚡ easy
Answer: From finite differences of first two boxes or zero velocity with high initial P—tradeoff between fast lock vs overshoot.
17 SORT’s use? 📊 medium
Answer: Each track maintains Kalman state; Hungarian matches detections to predicted boxes—simple, fast MOT baseline.
18 OpenCV? ⚡ easy
Answer: cv2.KalmanFilter with transition/measurement matrices—set dt, Q, R for bbox tracking experiments.
import cv2
kf = cv2.KalmanFilter(4, 2)  # example dims
19 Numerical issues? 🔥 hard
Answer: Use Joseph form for P update, symmetric enforcement, or square-root filtering if covariance becomes indefinite.
20 When Kalman fails? 📊 medium
Answer: Highly nonlinear motion, multi-modal uncertainty (occlusions), or heavy-tailed detector noise—consider particle, IMM, or learning-based motion.

Kalman Cheat Sheet

Loop
  • Predict
  • Update
Tuning
  • Q process
  • R measure
Nonlinear
  • EKF

💡 Pro tip: K trades prediction uncertainty vs measurement noise.

Full tutorial track

Go deeper with the matching tutorial chapter and code examples.