Visual Mesh Based Multi-Object Tracking
Closing Date: 31 March 2019
The project objective is to achieve the highest currently possible frame rates when tracking several targets of known size and shape from a static camera with high accuracy and robustly in high resolution video.
The Visual Mesh is a new super-fast deep convolutional neural network for object detection. It was invented at UON’s Robotics Lab in 2018 and first presented at the 2018 RoboCup Symposium (Houliston and Chalup, 2018). A new geometric input transform allows to fix a constant pixel sample density for the target objects independently of their distance from the camera if certain parameters are known, such as height of the camera over ground, lens specifications and the approximate size and geometry of the target objects. Given these constraints the Visual Mesh can track objects at very high frame rates and performs far superior to the currently best deep learning alternatives such as Yolo or SSD mobile net (Houliston and Chalup, 2018). So far the Visual Mesh has only been tested for single object tracking. The present project will study robustness and performance of the Visual Mesh when combined with real-world multi-object tracking (MOT) systems and develop a ready-to-use implementation with potential for commercialisation in defence relevant MOT. Methods for evaluating and monitoring tracking performance have to address tracking accuracy by determining false positives, false negatives and instances where the projections of object trajectories intersect and label switches occur. In addition to tracking accuracy, the precision of location in tracking has to be evaluated by calculating the intersections of the ground truth regions with the estimated regions of the segmentations or bounding boxes of the target objects. Traditional MOT systems have been based on state estimations using current and past information or, more sophisticated but slower approaches that also take future information through data association into account. One of the latter techniques is Markov-Chain Monte-Carlo Data Association that allows to deal with occlusions and the identification and removal of false positives. Deep learning based systems can utilise long short-term memory techniques that may integrate over time and space and allow to predict target trajectories and locations. The planned project will review and evaluate several of the best traditional MOT systems and then employ the Visual Mesh to transform some of them into a new high-performance system. Literature: T. Houliston and S. Chalup. Visual Mesh: Real-time Object Detection Using Constant Sample Density. RoboCup Symposium 2018, Springer LNCS, in press https://arxiv.org/abs/1807.08405
PhD Scholarship details
Funding: $27,082 per annum (2018 rate) indexed annually. For a PhD candidate, the living allowance scholarship is for 3.5 years and the tuition fee scholarship is for four years. An additional funding to top-up the baseline stipend and to contribute to running costs is also available.
Supervisor: Associate Professor Stephan Chalup
Available to: Australian Citizens
Applicants must be Australian Citizens and meet the eligibility criteria for admission.
Interested applicants should send an email expressing their interest along with scanned copies of their academic transcripts, CV, a brief statement of their research interests and a proposal that specifically links them to the research project.
Please send the email expressing interest to Stephan.Chalup@newcastle.edu.au by 5pm on 31 March 2019.
Applications Close 31 March 2019
|Contact||Associate Professor Stephan Chalup|
|Phone||+61 2 4921 6080|
PhD and Research MastersFind out more