No description available.

Conventional cameras are usually small in their field of view (FOV) and make the observable region limited. Applications by such a vision system may also limit motion capabilities for robots when it comes to object tracking. Omnidirectional camera has a wide FOV which can obtain environmental data from all directions. In comparison with conventional cameras, the wide FOV of omnidirectional cameras reduces blind regions and improves tracking ability. In this thesis, we assume an omnidirectional camera is mounted on a moving platform, which travels with planar motion. By applying optical flow and CAMShift algorithm to track an object which is non-propelled and only subjected to gravity. Then, by parabolic fitting, least-square method and Levenberg-Marquardt method to predict the 3D coordinate of the object at the current instant and the next instant, we can finally predict the position of the drop point and drive the moving platform to meet the object at the drop point. The tracking operation and drop point prediction can be successfully achieved even if the camera is under planar motion and rotation.

Robust object tracking plays a central role in many applications of image processing, computer vision and automatic control. In this thesis, robust object tracking under complex environments, including heavy clutters in the background, low resolution of the image sequences and non-stationary camera, has been studied. The interest of this study stems from the improvement of the performance of visual tracking using particle filtering. A Geometric Active contour-based Tracking Estimator, namely GATE, has been developed in order to tackle the problems in robust object tracking where the existence of multiple features or good object detection is not guaranteed. GATE combines particle filtering and the level set-based active contour method. The particle filtering method is able to deal with nonlinear and non-Gaussian recursive estimation problems, and the level set-based active contour method is capable of classifying state space of particle filtering under the methodology of one class classification. By integrating this classifier into the particle filtering, geometric information introduced by the shape prior and pose invariance of the tracked object in the level set-based active contour method can be utilised to prevent the particles corresponding to outlier measurements from being heavily reweighted. Hence, this procedure reshapes and refines the posterior distribution of the particle filtering. To verify the performance of GATE, the performance of the standard particle filter is compared with that of GATE. Since video sequences in different applications are usually captured by diverse devices, GATE and the standard particle filters with the identical initialisation are studied on image sequences captured by the handhold, stationary and PTZ camera, respectively. According to experimental results, even though a simple color observation model based on the Hue-Saturation-Value (HSV) color histogram is adopted, the newly developed. GATE significantly improves the performance of the particle filtering for object tracking in complex environments. Meanwhile, GATE initialises a novel approach to tackle the impoverishment problem for recursive Bayesian estimation using sampling method.

The aim of this thesis is to predict targets in a Multiple Object Tracking (MOT) task, in which subjects track multiple moving objects. We processed and analyzed data containing object and gaze position information from 1148 MOT trials completed by 20 subjects. We extracted multiple features from the raw data and designed a machine learning approach for the prediction of targets using neural networks and hidden Markov models. We assessed the performance of the models and features. The results of our experiments show that it is possible to train a machine learning model to predict targets with very high accuracy. 1

Tsunamis are among the most devastating and complex natural disasters, affecting coastal regions worldwide. Tsunami waves are generated through many natural phenomena, such as earthquakes, landslides, and volcanic eruptions. The waves travel at high speeds away from the source, potentially affecting multiple countries with very little warning. Over the past several decades, tsunamis such as the 2004 Indian Ocean, the 2010 Chilean, and the 2011 Tohoku Tsunami served as reminders of the potential devastation of these natural disasters, resulting in tragic loss of life and billions of dollars in damages. Forensic engineering field investigations and subsequent analysis of these events have demonstrated that infrastructure in these tsunami-prone regions was not adequately prepared for the extreme forces associated with a tsunami. As a result, there has been an increased research emphasis worldwide on the planning and design of infrastructure located in tsunami-prone areas to be better prepared for such future events. The present study aims to experimentally investigate and analyze the motion of debris carried by an inundating tsunami flood. One of the previous challenges involved in the evaluation of debris motion during such events was a lack of experimental methods that could non-invasively, quickly and accurately track the motion of debris at high velocities. This study introduces two innovative methods of tracking the debris. The first one used a novel camera-based tracking algorithm, while the second used Bluetooth and Inertial Measurement Unit sensors to track the debris within the inundating tsunami flood. The study outlines, for the first time, the technology and methods involved in the two tracking methods as it used both dry-test and wet-test experiments to evaluate the applicability of these methods in coastal and hydraulic engineering. This study used these two methods to evaluate the motion of debris from experiments conducted in a new Tsunami Wave Basin commissioned recently at Waseda University (Tokyo, Japan). The study examined the effect of the initial positioning of the debris, particularly focusing on the spreading area of the debris (determining thus their maximum displacement and the spreading angle of the debris). The results showed that an increase in the number of the debris resulted in an increase in the spreading angle of the debris and a decrease in the displacement of the debris. The increased number of debris also added significantly more variation in the final resting position of the debris due to the increased debris-debris collisions. The initial orientation of the debris also affected debris motion, particularly influencing the peak velocity of the debris and the distance from the initial debris resting position to where the peak velocity was observed.

Nowadays, researchers have shown convolutional neural network (CNN) can achieve an improved performance in multi-object tracking (MOT) by performing detection and re-identification (ReID) simultaneously. Many models have been created to overcome challenges and bring the state-of-the-art performance to a new level. However, due to the fact the CNN models only utilize feature from a local region, the potential of the model has not been fully utilized. The long range dependencies in spatial domain are usually difficult for a network to capture. Hence, how to obtain such dependencies has become the new focus in MOT field. One approach is to adopt the self-attention mechanism named transformer. Since it was successfully transferred from natural language processing to computer vision, many recent works have implemented it to their trackers. With the introduce of global information, the trackers become more robust and stable. There are also traditional methods which are re-designed in the manner of CNN and achieve satisfying performance such as optical flow. It can generate a correlated relation between feature maps and also obtain non-local information. However, the introduces of these mechanism usually causes a significant surge in computational power and memory. They also requires huge amount of epochs to train thus the training time is largely increased. To solve this issue, we propose a new method to gather non-local information based on the existing self-attention methods, we named it dual attention with regional-representation, which significantly reduces the training time as well as the inference time, but only causes a small increase in computational memory and are able to run with a reasonable speed. Our experiments shows this module can help the ReID be more stable to improve the performance in different tasks. / Thesis / Master of Applied Science (MASc)

Effective multi-object tracking is crucial for autonomous vehicles to navigate safely and efficiently in dynamic environments. To make autonomous vehicles more affordable one area to address is the computational limitations of the sensors, therefore, cameras are often the first choice sensor. Three challenges in implementation of multi-object tracking in autonomous vehicles are: 1) In these vehicles, sensors like cameras are not static, which can cause motion blur in the frames and make tracking inefficient. 2) Traditional methods for motion compensation, such as those used in Kalman Filter-based Multi-Object Tracking, require extensive parameter tuning to match features between consecutive frames accurately. 3) Simple intersection over union (IoU) metric is insufficient for reliable identification in such environments. This thesis proposes a novel methodology for 2D multi-object tracking in autonomous vehicles using a camera-based Tracking-by-Detection (TBD) approach, emphasizing four key innovations: (1) A real-time deblurring module to mitigate motion blur, ensuring clearer frames for accurate detection; (2) deep learning-based motion compensation module that adapts dynamically to varying motion patterns, enhancing robustness; (3) adaptive cost function for association, incorporating object appearance and temporal consistency to improve upon traditional IoU metrics; (4) The integration of the Unscented Kalman Filter to effectively address non-linearities in the tracking process, enhancing state estimation accuracy. To maintain a Simple Online and Realtime (SORT) framework, we enhance detection by fine-tuning YOLOv8 and YOLOv9 models using autonomous driving datasets like BDD100K and KITTI, which are specifically tailored for these scenarios. Additionally, we incorporate a non-linear approach using the UKF to better capture the influence of various tracking dynamics, further improving tracking performance. Our evaluations show that the proposed methodology significantly outperforms existing state-of-the-art methods while maintaining the same inference rate as the baseline SORT model. These advancements not only improve the accuracy and reliability of multi-object tracking but also reduce the computational burden associated with parameter tuning and motion compensation. Consequently, this work presents a robust and efficient tracking solution for autonomous vehicles, making it viable for real-world deployment under both computational and cost constraints. / Master of Science / Tracking multiple objects is really important for self-driving cars to move safely in busy places. Cameras are often the best choice because they are cheaper and easier to use, but using cameras comes with three main challenges: (1) When cars move, cameras can make blurry images, which makes it harder to see and track things; (2) Traditional tracking methods, like Kalman Filters, need a lot of adjustments to work well; (3) Simple methods, like checking if objects overlap (called Intersection over Union), are not always good enough in crowded, complicated places. This thesis presents a new way to track lots of things using cameras, with four big improvements: (1) A real-time deblurring system that fixes blurry pictures so the camera can see things more clearly; (2) A smart system that uses deep learning to follow movement better; (3) A better way to match objects by using not just their positions but also how they look and move over time, which is better than old IoU methods; (4) A special tool called the Unscented Kalman Filter that helps track objects more accurately when their movements aren't simple or straight. To keep everything simple, fast, and real-time, we use object detectors to help find objects, and we train them with special self-driving datasets like BDD100K and KITTI. These datasets are great for showing the kinds of situations self-driving cars deal with. The Unscented Kalman Filter helps us track objects with more complicated movements, making everything more accurate. Our study show that this new way works much better than older methods, without making the system slower. These improvements make tracking more reliable and cut down on the time needed for tuning and adjusting. Overall, this work provides a strong and simple solution for tracking things in self-driving cars, even if the computer isn't super powerful or the budget is small.

Multiple object tracking is the process of assigning unique and consistent identities to objects throughout a video sequence. A popular approach to multiple object tracking, and object tracking in general, is to use a method called tracking-by-detection. Tracking-by-detection is a two-stage procedure: an object detection algorithm first detects objects in a frame, these objects are then associated with already tracked objects by a tracking algorithm. One of the main concerns of this thesis is to investigate how different object detection algorithms perform on surveillance video supplied by National Forensic Centre. The thesis then goes on to explore how the stand-alone alone performance of the object detection algorithm correlates with overall performance of a tracking-by-detection system. Finally, the thesis investigates how the use of visual descriptors in the tracking stage of a tracking-by-detection system effects performance.  Results presented in this thesis suggest that the capacity of the object detection algorithm is highly indicative of the overall performance of the tracking-by-detection system. Further, this thesis also shows how the use of visual descriptors in the tracking stage can reduce the number of identity switches and thereby increase performance of the whole system.

In this thesis we directly consider an object tracking problem for wireless sensor networks (WSNs), called track persistence. Track persistence temporally extends the problem of object tracking by seeking to store and retrieve the entire history of an object. To provide an initial solution to track persistence, we develop two distinct algorithms. The first algorithm, update to sink, translates track persistence into a centralized problem. The second algorithm, a linked list-like algorithm, builds a dynamic data structure as the object traverses the network, and rebuilds the object history distributively upon demand. We conduct worst case analysis upon both of these algorithms. Finally, we implement a simulation environment and run a number of tests upon both algorithms. Track persistence is a very challenging problem, and this thesis contributes a pair of solutions which stand as a basis for future research. / Thesis (Master, Computing) -- Queen's University, 2010-09-09 12:56:50.921

No description available.