\prefacesection{Abstract} Multi-object tracking (MOT) is a key building block in video analytics and finds extensive use in surveillance, search and rescue, and autonomous driving applications. Object detection, a crucial stage in MOT, dominates in the overall tracking inference time due to its reliance on Deep Neural Networks (DNNs). Despite the superior performance of cutting-edge object detectors, their extensive computational demands limit their real-time application on embedded devices that possess constrained processing capabilities. Hence, we aim to reduce the computational burdens of object detection while maintaining tracking performance. As the first approach, we adapt frame resolutions to reduce computational complexity. During inference, frame resolutions can be tuned according to the complexity of visual scenes. We present DeepScale, a model-agnostic frame resolution selection approach that operates on top of existing fully convolutional network-based trackers. By analyzing the effect of frame resolution on detection performance, DeepScale strikes good trade-offs between detection accuracy and processing speed by adapting frame resolutions on-the-fly. Our second approach focuses on enhancing the efficiency of a tracker by model adaptation. We introduce AttTrack to expedite tracking by interleaving the execution of object detectors of different model sizes in inference. A sophisticated network (teacher) runs for keyframes only while, for non-keyframe, knowledge is transferred from the teacher to a smaller network (student) to improve the latter’s performance. Our third contribution involves exploiting temporal-spatial redundancies to enable real-time multi-camera tracking. We propose the MVSparse pipeline which consists of a central processing unit that aggregates information from multiple cameras (on an edge server or in the cloud) and distributed lightweight Reinforcement Learning (RL) agents running on individual cameras that predict the informative blocks in the current frame based on past frames on the same camera and detection results from other cameras. / Thesis / Doctor of Science (PhD)

Background.Identifying the quality of the products in the manufacturing industry is a challenging task. Manufacturers use needles to print unique numbers on the products to differentiate between good and bad quality products. However, identi- fying these needle printed characters can be difficult. Hence, new technologies like deep learning and optical character recognition (OCR) are used to identify these characters. Objective.The primary ob jective of this thesis is to identify the needle-printed characters on steel bars. This ob jective is divided into two sub-ob jectives. The first sub-ob jective is to identify the region of interest on the steel bars and extract it from the images. The second sub-ob jective is to identify the characters on the steel bars from the extracted images. The YOLOV5 and YOLOV5-obb ob ject detection algorithms are used to achieve these ob jectives. Method. Literature review was performed at first to select the algorithms, then the research was to collect the dataset, which was provided by OVAKO. The dataset included 1000 old images and 3000 new images of steel bars. To answer the RQ2, at first existing OCR techniques were used on the old images which had low accuracy levels. So, the YOLOV5 algorithm was used on old images to detect the region of interest. Different rotation techniques are applied to the cropped images(cropped after the bounding box is detected) no promising result is observed so YOLOV5 at the character level is used in identifying the characters, the results are unsatisfactory. To achieve this, YOLOV5-obb was used on the new images, which resulted in good accuracy levels. Results. Accuracy and mAP are used to assess the performance of OCRs and selected ob ject detection algorithms. The current study proved Existing OCR was also used in the extraction, however, it had an accuracy of 0%, which implies it failed to identify characters. With a mAP of 0.95, YOLOV5 is good at extracting cropped images but fails to identify the characters. When YOLOV5-obb is used for attaining orientation, it achieves a mAP of 0.93. Due to time constraint, the last part of the thesis was not implemented. Conclusion. The present research employed YOLOV5 and YOLOV5-obb ob ject detection algorithms to identify needle-printed characters on steel bars. By first se- lecting the region of interest and then extracting images, the study ob jectives were met. Finally, character-level identification was performed on the old images using the YOLOV5 technique and on the new images using the YOLOV5-obb algorithm, with promising results

Autonomous vehicles require intelligent systems to perceive and navigate unstructured envi- ronments. The scope of this project is to improve and develop algorithms and methods to support autonomy in the off-road problem space. This work explores computer vision architectures to support real-time object detection. Furthermore, this project explores multimodal deep fusion and sensor processing for off-road object detection. The networks are compared to and based off of the SqueezeSeg architecture. The MAVS simulator was utilized for data collection and semantic ground truth. The results indicate improvements from the SqueezeSeg performance metrics.

<p> Body Posture provides enough information regarding the current state of mind of a person. This idea is used to implement a system that provides feedback to lecturers on how engaging the class has been by identifying the attentive levels of students. This is carried out using the posture information extracted with the help of Mediapipe. A novel method of extracting features are from the key points returned by Mediapipe is proposed. Geometric moments aided features classification performs better than the general distances and angles features classification. In order to extend the single person pose classification to multi person pose classification object detection is implemented. Feedback is generated regarding the entire lecture and provided as the output of the system. </p>

No description available.

No description available.

No description available.

No description available.

This project aims to create a robot system that can accurately figure out if the throws are successful. This can help make various industrial tasks more efficient. The system uses implemented methods to process data from fisheye camera data and depth sensor data, to check the quality of the throws. The main goal is to find out if the thrown object reaches its target or not, with more advanced tasks including predicting its path when frames are lost or not tracked properly.To put the system together the Robot Operating System (ROS) was used for handling data and processing, as well as different tools and techniques, like bag files and OpenCV. A variety of methods and algorithms were used to apply background subtraction, clustering, curve fitting, marking objects and drawing the path they take in the air. The depth sensor data processing is included to make up for the limitations of 2D camera data, providing more accurate and reliable tracking of thrown objects.

京都大学 / 新制・課程博士 / 博士(農学) / 甲第24913号 / 農博第2576号 / 新制||農||1103(附属図書館) / 京都大学大学院農学研究科地域環境科学専攻 / (主査)教授 飯田 訓久, 教授 近藤 直, 教授 野口 良造 / 学位規則第4条第1項該当 / Doctor of Agricultural Science / Kyoto University / DFAM