Browsing by Author "Buckton, Calib Jonas"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemMulti-spectral object tracking and prediction of kinematic quantities(Stellenbosch : Stellenbosch University, 2021-12) Buckton, Calib Jonas; Wyngaardt, Shaun M.; Malaza, Vusi D.; Stellenbosch University. Faculty of Science. Dept. of Physics.ENGLISH ABSTRACT: The use of convolutional neural networks in object identification is well-documented and useful in a variety of applications. Particularly, when combined with a tracking algorithm such as SORT or DeepSORT, a convolutional network can lay the foundation for accurately tracking multiple objects in motion. There are many well-known cases of observation of motion leading to hypothesis and modelling through experimentation, particularly in object motion and trajectory. However, this work investigates the viability of deep learning models and recursive filtering to perform observations and predictions based on models. This leads to the idea of using neural networks and tracking algorithms to perform observation and tracking of an object’s trajectory. In addition, there is the extension of this application to motion prediction, such as estimating the future trajectory of an object, which can further be used to determine other useful kinematic quantities. For tracking an object in motion, an object tracking algorithm will be used to track moving vehicles. Generally, this can be applied to track any object. The use of Kalman filters is common for these estimation tasks, particularly for estimating an object’s position at a time dt a few seconds ahead. With the use of a suitable motion model for the objects being tracked, one can increase the effectiveness of these filters. Another compelling idea is the combination of multiple Kalman filters into one estimator, since a single filter can hold just one motion model. Each filter, in this way, can account for at least one possible state of the object in motion. An attempt is also made to apply this algorithm to a non-visible spectrum. The infrared spectrum is useful for tracking in low-light environments, or for tracking thermal information. This can be achieved by applying an infrared filter, or the use of suitable infrared camera to obtain an infrared dataset. Finally, there is an interacting multiple model estimator for predicting future states of objects being tracked. Such a filter is composed primarily of Kalman filters, each with their own motion model. A comparison is made between ground truth trajectory data and predictions from the estimator. Choice of coordinate system can also be important when tracking objects in motion. For example, for application to real-time GPS logging, these coordinates will need to be sensible enough for estimation outside of the camera pixel boundaries. For this problem, note the usefulness of a perspective mapping, using homogeneous coordinates and relevant GPS data [1].