Multi-target object tracking is a challenging computer-vision problem with a diverse range of applications ranging from security, to sports-analysis, and animal-welfare. This problem involves recording the position all objects of interest, such as people or animals, at each time-step. This is a very challenging task due to complex appearance changes, long-term occlusions, and the difficulty of accurately detecting deformable objects from multiple viewpoints. The problem of multi-target tracking is still largely unsolved for the general case. Most approaches to date consist of a complex pipeline of different software to detect objects, extract appearance models, build a graph of evidence, associate detections into short tracklets, then associate tracklets into longer tracks. This software pipeline is computationally inefficient and makes it hard to improve tracker performance as so many different components must be considered.
Taking inspiration from other areas of computer-vision, where end-to-end trained deep-learning method have proliferated, the field of multi-target tracking has recently started to move towards learning-based approaches. These include regression-based trackers that take into account additional image information outside object detections [1,2], and the use of graph networks to better integrate all of the evidence needed to perform tracking [3-7].
In this PhD project we propose to extend and adapt the recent learning-based approaches to multi-target tracking to the animal welfare use-case. This will require adapting and extending existing methods to cope with the unique challenges of this domain. The eventual aim would be to create a complete real-time end-to-end tracking solution that could be deployed with minimal human intervention.
Some problems specific to multi-target tracking for the animal welfare domain include: the fact that the appearance of animals may be very similar, meaning that other sources of evidence, such as motion of individuals or very fine-grained appearance differences need to be modelled. There is also the problem of performing tracking in scenes that have heterogeneous motion and crowd density, such as can be observed in many farm environments. This means that some individuals will move and behave very differently from their neighbours and that naïve tracking by detection methods begin to break down due to the different ques need to detect individuals in sparse or dense environments.
To deal with these issues we would like to build a regression-based tracker that can integrate low-level pixel information from multiple frames in order to produce precise tracklets in the presence of detector inaccuracy. Additionally, we would like to apply recent work in graph networks to the tracking problem. A deep graph network would be used to combine spatial and temporal information to predict the association of objects. Recent advances in attention-based architectures would be used to better focus the tracker’s attention on relevant evidence while ignoring distractors. As a further step, we also propose to integrate the appearance extraction and object detection components into the evidence available to the graph network so that all relevant information can be combined to produce accurate tracking decisions. By integrating all components into a single end-to-end solution, it should be possible to improve efficiency, by reducing repeated computation and complex tracking pipelines, so that real time operation is achieved.
Objectives:
- Build a regression-based tracker that combines appearance and motion information across frames to create accurate low-level tracklets
- Investigate use of graph networks for data association. Investigate how prior knowledge of motion models can be taken into account, and how attention mechanisms can be used to focus network’s attention on relevant information
- Integrate regression-tracker with graph-network-based data-association so that object detection and appearance models networks can be trained end-to-end with data association
Academic Requirements:
A minimum 2.1 honours degree or equivalent in Computer Science or Electrical and Electronic Engineering or relevant degree is required.
Applicants should apply electronically
through the Queen’s online application portal at: https://dap.qub.ac.uk/portal/
Further information available at: https://www.qub.ac.uk/schools/eeecs/Research/PhDStudy/
Closing date for applications for is Monday 22nd March
Continue with Facebook
