Object Tracking

We mainly put our emphasis on autonomtive object tracking, especially the extended object tracking.

Tracker uses mainly three elementary knowledge:

  • noise model: uncertainties.

  • system dynamics: state evolution and prior information,

  • measurement system: sensor,

to extract two essential information:

  • quality, i.e., accuracy and reliability, higher than the raw measurements

  • inference information, not directly available in the measurements.

For serious Bayesian researchers, it is highly recommended to read [1]:

  1. Purple Bible:

    Estimation with Applications to Tracking and Navigation by Yaakov Bar-Shalom, X. Rong Li and Thiagalingam Kirubarajan.

  2. Yellow Brick:

    Tracking and Data Fusion: A Handbook of Algorithms by Yaakov Bar-Shalom, Peter K. Willett and Xin Tian.

  3. Grey Manual:

    Design and Analysis of Modern Tracking Systems (Artech House Radar Library) by Samuel Blackman and Robert Popoli.

Stone Soup by DSTL

A wonderful point target tutorial for industrial engineers is presented by the Defence Science and Technology Laboratory (DSTL), called Stone Soup.

Stone Soup is developed in Python with six major components: framework, data, algorithms, metrics, simulators, and sensor models. The framework is the core of the project and as a software architecture in a modular fashion integrating all essence in Tracking (such as dynamic/measurement models, noise metrics and simulators).

Stone Soup Structure

Fig. 01: Stone Soup Structure

One of our major efforts is to design an Extended Object Tracking Framework similar to the Stone Soup.

However, before applying the tracking/estimation theory originated from aerospace systems for automotive applications, we must understand their common ground and differences.

Automotive Tracking versus Aerospace Tracking

Automotive

Aerospace

weak Gaussian assumption

strong Gaussian assumption

extended object

point target

kinematic and shape association

probablistic association

Let us call the system OCEAN.

Ocean by UniverSee

The Ocean consists mainly of the following modules:

  1. Framework for extended object tracking (EOT)

  2. Dynamic model set

  3. Measurement model set

  4. Mutltiple model toolbox

  5. Association scheme

  6. Performance metrics

  7. Pre-Processing tool