Projects and Codes

Multi-target filtering

This project focuses on multi-target filtering using Bayesian random finite sets and recurrent neural networks. The algorithms are based on the following papers:

M. Emambakhsh, A. Bay, and E. Vazquez, “Filtering Point Targets via Online Learning of Motion Models,” https://arxiv.org/abs/1902.07630, 2019. PDF

M. Emambakhsh, A. Bay, and E. Vazquez, "Deep Recurrent Neural Network for Multi-target Filtering", 25th International Conference, MMM 2019, Thessaloniki, Greece, pp. 519 -531, 2018. PDF Presentation Video

The results are obtained from two sequence of datasets, a synthetic and real data as shown below. The clutter is shown as black dots, while the detected targets and ground truth are illustrated as red and green, respectively.

Multi-target filtering over synthetic data.

Multi-target filtering over real data.

Multi-target tracking

In this work we propose an algorithm which learns the motion model for multi-target tracking and filtering:

M. Emambakhsh, A. Bay, and E. Vazquez, "Convolutional Recurrent Predictor: Implicit Representation for Multi-target Filtering and Tracking", IEEE Transactions on Signal Processing, vol. 67, no. 17, pp. 4545-4555, 2019. PDF

The algorithm is evaluated over several datasets, including the Multi-Object Tracking (MOT) dataset. The results on one sequence from MOT17 is shown below.

Multi-sensor simulation for a moving vehicle.

The developed algorithm is then applied to a real world scenario, to perform sensor fusion at data level, over radar, Lidar and stereo camera. The stereo camera and bird's-eye view of the fusion results are shown in the following figures.

(a)

(b)

(a) The view from the stereo camera of a scene occupied by pedestrians and road obstacles: a typical test scene for a driver-less car; (b) A bird's-eye view of the fusion results over the ground plane.

Nose as a biometric

This is an implementation of the feature extraction step explained in the following paper using spherical patches over the nasal region: CODE!

Although it has been introduced for 3D face recognition using the nasal region, its applications can also be extended to any classification task and modaality such as 2D, 2.5D or 3D images.

M. Emambakhsh and A. Evans, “Nasal patches and curves for an expression-robust 3D face recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), vol. 39, no. 5, pp. 995-1007, 2017.

Input nasal region and a set of landmarks.

Maximal Gabor-wavelet outputs along different orientations for each scale.

Normal maps computed per scale.

The spherical patches over the nasal region, used as feature descriptors.

Final extracted feature vector.

Medical image segmentation via neural networks

This project aimed at medical image segmenting tested over MRI brain images: CODE!

M. Emambakhsh and M. Sedaaghi, “Automatic MRI brain segmentation using local features, self-organizing maps, and watershed,” in IEEE International Conference on Signal and Image Processing Applications (ICSIPA 2009), Kuala Lumpur, Malaysia, 2009, pp. 129–134.

First, three sets of features are extracted from the MRI image:

Gaussian filtered

Mediated

Morphological filtered

Then after initial dimensionality reduction using PCA, SOM networks are used in two steps for feature mapping:

An example feature map of the first layer

The network is used to cluster the input image:

Finally, the edge map created from the clustering algorithm is fed into a watershed segmentation algorithm:

Edge map

Segmentation results

Learning to do approximate computing

In this project, we proposed an approach to incorporate the prior knowledge about the problem at hand, in order to perform approximate computing. Our method dynamically modifies the approximation level at run time.

The algorithm is evaluated over an extended Kalman tracking problem, in which the target has a circular displacement and the sensor is located at the origin detecting the target's bearing and range.

The results over this toy problem shows 2.54% energy reduction while only 0.1661 (unit distance) tracking error is obtained:

The following video gives a short presentation of this work:

P. Garcia, M. Emambakhsh, and A. Wallace, “Learning to approximate computing at run-time,” IET 3rd International Conference on Intelligent Signal Processing (ISP 2017), London, UK, 2017, pp.1-6. PDF Code

EKF SLAM

This is a GUI for the extended Kalman filter (EKF) simultaneous localisation and mapping (SLAM), with known associations explained in: page 314 of the Probabilistic Robotics book by Sebastian Thrun, Wolfram Burgard and Dieter Fox: CODE!

Once a target (landmark) is detected, it is localised over the map on the right hand side. At the same time the vehicle's location is updated using the EKF SLAM framework. When a target is in the vehicle's FoV, it is coloured as blue. The targets and vehicle's states at any time is shown the middle column at the top and bottom, respectively.

The uncertainty of the landmarks' locations and vehicle's localisation gradually decreases over time.