Mastering Target Tracking for ADAS / Autonomous Driving

Explain conditional probability and the formula p(a|m) for events a and m, highlighting prior information. Illustrate data reduction with seasons and summer and apply p(a|m) properties to probability calculations.

Apply Bayes' theorem to update the probability of A as new information about related events becomes available, refining P(A|B) using P(B|A) and P(A) with P(B) in the denominator.

Explore how autonomous driving systems detect obstacles using LiDAR, radar, camera, and ultrasonic sensors, and update obstacle probabilities with total probability theorem and Bayes theorem after new radar measurements.

Define random experiments, perform trials, map outcomes to events, and apply unconditional probability, conditional probability, total probability theorem, and Bayes' theorem to discrete events and real-line density and distribution functions.

Explore random variables, probability density and distribution functions, and how to compute them, then study conditional probability, total probability theorem, joint distributions, and defining random variables as functions of others.

Define X as stopping distance and Y as reaction time, random variables, then form Z as total distance from X and Y, and compute Z's distribution via joint pdf integral.

Define a random experiment and map outcomes to a random variable. Apply probability density, distribution, conditional probability, total probability theorem, Bayes theorem to derive statistics and joint distributions.

Explore conditional expectation in an emergency brake assist, showing how Y's braking distance is the conditional expectation given X, and how the unconditional expectation equals the outer expectation.

Explore joint moments for two random variables, deriving covariance and correlation from a two-dimensional Gaussian joint density to assess sensor mounting errors on orientation estimates in autonomous driving.

Study random vectors and random processes, including the law of large numbers and Markov, stationary, and ergodic processes, with related statistical metrics for ADAS tracking.

Explore the law of large numbers, comparing weak and strong forms, as sample averages converge to the true mean for independent, identically distributed random variables, with a lightbulb quality-control example.

Define a random process formally as a collection of random variables indexed by time or space, mapping time and outcomes to a real value, with joint probability densities across time.

Break down the joint probability density into conditional densities and show that a Markov process lets Xn depend only on Xn-1, simplifying the joint density into a product of conditionals.

Learn how ergodic processes let a single long observation reveal a system's statistics by equating time averages with ensemble averages, illustrated with autonomous car sensor noise.

Explore signals and systems, contrasting probabilistic and deterministic foundations for tracking algorithms. Examine continuous and discrete time systems, memory, causality, linearity, time invariance, observability, Fourier transform, impulse response, and convolution.

Explore how a system responds to a sudden input by examining impulse response, the output to a Dirac delta, revealing stability, causality, and frequency response.

Explore the frequency domain by decomposing signals into constituent frequencies with the Fourier transform, contrasting time-domain intuition and revealing amplitude, phase, and spectra for filtering and analysis.

Apply stationary random processes to a linear time-invariant system to obtain the output via convolution with impulse response, and explore continuous and discrete time with autocorrelation and power spectral density.

Predictor-corrector techniques are iterative methods in numerical analysis that predict a solution to differential equations and refine it with observed data. Meteorologists apply this approach to adjust forecasts.

Categorize estimation by what is estimated, the algorithm, and the problem; cover state, parameter, signal, function, hypothesis, and decision estimations in ADAS with linear, dynamic, and stochastic contexts.

Explore state estimation for autonomous driving by turning noisy sensor measurements into filtered ego and target states, deriving properties like target type and road context.

Combine lidar, ultrasonic, and radar data with a minimum mean square estimator to improve target distance accuracy. Derive optimal coefficients from cross- and auto-correlations for sensor fusion in autonomous driving.

Use the least squares estimator to fit a linear model relating wheel slip and braking force to acceleration, accounting for road friction, with offline calibration and recursive, sliding-window real-time updates.

Compare MMSE and LSE, highlighting when random versus deterministic inputs apply, recursive and nonlinear forms, and when generalized or weighted least squares are needed.

Explore how gradient descent drives adaptive algorithms to minimize a performance criterion by updating coefficients step by step. Understand how the gain, gradient direction, and potential local minima influence convergence.

Explore standard Kalman filter analysis by showing how range estimates converge to the actual value, stabilize noise, and form a stationary process with a known standard deviation.

Extend the Kalman filter to handle nonlinear dynamics and measurements by linearizing with jacobians, recovering Fk and Hk to enable accurate target tracking.

Learn how an extended Kalman filter tracks a crossing vehicle using radar azimuth measurements in a constant-velocity ADAS scenario, deriving the state estimate of x, y, vx, vy.

Apply an extended Kalman filter to track parallel cars with radar by incorporating azimuth and range measurements, updating the two-row measurement matrix, and executing prediction and update steps.

Demonstrates an extended Kalman filter for parallel cars using an EKF simulator dashboard, tuning azimuth and range noise and observing convergence, accuracy within 0.5 meters for driving scenarios.

Learn how to decouple range measurements in an EKF for overtaking lane changes, using range x and range y instead of azimuth and Doppler, updating the H matrix before filtering.

Explore an EKF simulator dashboard that models overtaking ego in lane changes with decoupled X and Y tracking for maneuvering targets, achieving within 0.5 meters error.

Wrap up your learning on mastering target tracking for multiple sensors for ADAS and autonomous vehicles, and share feedback and future topic requests in the comments.