Multi-target filtering: definition & applications
● A multi-target filtering
algorithm removes clutter
(false positive) from a
sequential data.
● Examples:

Stereo vision

Radar/LiDAR signal analysis

Robotics: SLAM and occupancy grid

Object detection and tracking
● Example of multi-target filtering usage: varying
detection threshold to increase TPR

Multi-target filtering: algorithms
● Kalman Filter (KF) based techniques:

Gaussian and linear motion model assumptions.

Extended KF: Linearisation via Taylor series expansion to maintain Gaussian behaviour

Unscented KF: Deterministic selection of sample at different variances along each dimension to compute
covariance and mean for an estimated Gaussian function.
● Information Filter (IF)

Similar to KF, but works on canonical space (inverted covariance, i.e. information matrix)

SEIF: unlike the covariance matrix, the information matrix is very sparse. SEIF uses this sparsity to discard
significant number of landmarks at each iteration, improving computation time.
● Particle filter

Estimates the posterior using Monte Carlo technique. Can handle non-linear non-Gaussian models.
● Neural networks:

Recurrent neural networks

Long short-term memory

Multi-target filtering: challenges
● Extension to multi-target:

Random finite sets (RFS) and Probability Hypothesis Density (PHD)

Mapping the multi-target state vectors to a universal single target problem
● Fixed motion model issues:

Complex non-linear motion (can happen in presence of a noisy detector) can lead to wrong predictions.
● Gaussian assumption, especially in KF-PHD
● Challenges in using sequential learning algorithms:

Unlike the Bayesian generative models, they rely on a separate train/test steps

Cluttered unlabelled data can lead to weak predictive models

Variable input size

Memory management

Proposed multi-target filtering algorithm

Proposed algorithm
● The proposed algorithm addresses the following problems:

Handing non-linear non-Gaussian multi-target motion

Does not rely on a fixed motion model and it learns it incrementally

Use of neural networks (an example of a sequential learning algorithm) for filtering

Proposed algorithm: prediction step
● A target tuple is defined as:
● The model initially is trained to act as an auto-encoder regressor.
● Once the model is trained, it is transferred to the other target, saving time and memory.

Proposed algorithm: data association & filtering
● Using the incoming measurement RFS a set of residual tuples are computed:

Proposed algorithm: data association & filtering
● Computing all the ‘targetness’ error T for all measurements and targets creates a matrix:

Proposed algorithm: Update
● Using the targetness matrix:
●
●
●
●
●
●
●
●
● Then the following data association
algorithm is used assign target survival
(true positivity), death (false positivity)
and birth →

Proposed algorithm: Complexity analysis
● A basic Hungarian Matching:
● GM-PHD filter:
● The proposed method:

Multi-target filtering: experimental results
● The proposed algorithm is applied to a
synthetic multi-target filtering scenario:

Multiple scenarios are considered, such
as:

Variable number of targets

Non-linear motion

Birth/spawn of targets

Dense random clutter with a Poisson
distribution

Noisy measurement

Occlusion

Merge of targets ● Temporal overlay
visualisation of the
filtering result

● Optimal Sub-Pattern Assignment (OSPA) is used as the quantitative metric.

● Robustness against clutter density:

Conclusions and future work
● An algorithm is proposed to address non-linearity and fixed motion model challenges of the
available multi-target filtering algorithms.
● It is based on the use of a novel target tuple definition, LSTM architecture for motion
modelling and a linearly complex data association step.
● Future work:

Real data

Other applications: tracking, detection, etc.

End to end implementation

Thank you
www.cortexica.com
3rd Floor – 30 Stamford Street
WeWork Southbank Central London
SE1 9LQ
+44 (0) 203 868 8880
info@cortexica.com
Twitter: @cortexica

Deep Recurrent Neural Network for Multi-target Filtering

More Related Content

What's hot (10)

Similar to Deep Recurrent Neural Network for Multi-target Filtering (20)

More from Mehryar (Mike) E., Ph.D. (7)

Recently uploaded (20)

Deep Recurrent Neural Network for Multi-target Filtering