No description available.

Detta examensarbete har utförts under tio veckor på uppdrag av företaget BitSim NOW. Den manuella metod som idag används för mätning av stötar inom kulstötning kan utgöra en risk för felaktiga resultat och personskador. Med hjälp av tekniska hjälpmedel kan en lösning med noggrannare mätningar och lägre risk för skador implementeras i sporten kulstötning. Denna rapport presenterar en lösning som med hjälp av artificiell intelligens identifierar kulan utifrån en filmsekvens. Därefter beräknas längden av stöten med hjälp av en formel för kastparabeln. Lösningen jämförs sedan med en metod utan artificiell intelligens för att fastställa den bästa av de två metoderna. De variablersom jämfördes var noggrannheten på stötens längd och hur bra de två olika metoderna spårade kulan. Resultatet analyserades i relation till de uppsatta målen och sattes därefter in i ett större sammanhang. / This thesis project has been done during ten weeks on behalf of the companyBitSim NOW. The current method used to measure the length of shot-puts presents a risk of inaccurate results along with the risk of injury for the measuring personnel. With the help of technical aids, a solution with more accurate measurements and a lower risk for injuries could be implemented in the sport of shot-puts. This report presents a solution using artificial intelligence to first identify the shotin video films and secondly calculate the length using mathematical formulas. Thesolution is then compared to a method that does not use artificial intelligence, to determine what method is the superior one. The parameters that were compared were the accuracy of the length and the quality of the tracking. The result was analyzed in relation to the aims of the project and then put into a larger context.

Available computing resources play a large part in enabling the training of modern deep neural networks to complete complex computer vision tasks. Improving the efficiency with which this computational power is utilized is highly important for enterprises to improve their networks rapidly. The first few training iterations over the data set often result in substantial gradients from seeing the samples and quick improvements in the network. At later stages, most of the training time is spent on samples that produce tiny gradient updates and are already properly handled. To make neural network training more efficient, researchers have used methods that give more attention to the samples that still produce relatively large gradient updates for the network. The methods used are called ''Importance Sampling''. When used, it reduces the variance in sampling and concentrates the training on the more informative examples. This thesis contributes to the studies on importance sampling by investigating its effectiveness in different contexts. In comparison to other studies, we more extensively examine image classification by exploring different network architectures over a wide range of parameter counts. Similar to earlier studies, we apply several ways of doing importance sampling across several datasets. While most previous research on importance sampling strategies applies it to image classification, our research aims at generalizing the results by applying it to object detection problems on top of image classification. Our research on image classification tasks conclusively suggests that importance sampling can speed up the training of deep neural networks. When performance in convergence is the vital metric, our importance sampling methods show mixed results. For the object detection tasks, preliminary experiments have been conducted. However, the findings lack enough data to demonstrate the effectiveness of importance sampling in object detection conclusively.

In this master thesis, different neural networks have investigated annotating objects in video streams with partially annotated data as input. Annotation in this thesis is referring to bounding boxes around the targeted objects. Two different methods have been used ROLO and GOTURN, object detection with tracking respective object tracking with pixels. The data set used for validation is surveillance footage consists of varying image resolution, image size and sequence length. Modifications of the original models have been executed to fit the test data.  Promising results for modified GOTURN were shown, where the partially annotated data was used as assistance in tracking. The model is robust and provides sufficiently accurate object detections for practical use. With the new model, human resources for image annotation can be reduced by at least half. / I detta examensarbete har olika neurala nätverk undersökts för att annotera objekt i videoströmmar med partiellt annoterade data som indata. Annotering i denna uppsats syftar på avgränsninglådor runt de eftertraktade objekten. Två olika metoder har använts ROLO och GOTURN, objektdetektering med spårning respektive objektspårning av pixlar. Datasetet som användes för validering är videoströmmar från övervakningskameror i varierande bildupplösning, bildstorlek och sekvenslängd. Modifieringar av ursprungsmodellerna har utförts för att anpassa testdatat. Lovande resultat för modifierade GOTURN visades, där den partiella annoterade datan användes som assistans vid spårning. Modellen är robust och ger tillräckligt noggranna objektdetektioner för praktiskt bruk. Med den nya modellen kan mänskliga resurser för bild annotering reduceras med minst hälften.

Systems like autonomous vehicles may require real time embedded image processing under hardware constraints. This paper provides directions to design time and resource efficient Haar cascade detection algorithms. It also reviews some software architecture and hardware aspects. The considered algorithms were meant to be run on platforms equipped with a CPU and a GPU under power consumption limitations. The main aim of the project was to design and develop real time underwater object detection algorithms. However the concepts that are presented in this paper are generic and can be applied to other domains where object detection is required, face detection for instance. The results show how the solutions outperform OpenCV cascade detector in terms of execution time while having the same accuracy. / System så som autonoma vehiklar kan kräva inbyggd bildbehandling i realtid under hårdvarubegränsningar. Denna uppsats tillhandahåller anvisningar för att designa tidsoch resurseffektiva Haar-kasad detekterande algoritmer. Dessutom granskas en del mjukvaruarkitektur och hårdvaruaspekter. De avsedda algoritmerna är menade att användas på plattformar försedda med en CPU och en GPU under begränsad energitillgång. Det huvudsakliga målet med projektet var att designa och utveckla realtidsalgoritmer för detektering av objekt under vatten. Dock är koncepten som presenteras i arbetet generiska och kan appliceras på andra domäner där objektdetektering kan behövas, till exempel vid detektering av ansikten. Resultaten visar hur lösningarna överträffar OpenCVs kaskaddetektor beträffande exekutionstid och med samtidig lika stor träffsäkerhet.

This project aims at using deep learning to solve the pedestrian tracking problem for Autonomous driving usage. The research area is in the domain of computer vision and deep learning. Multi-Object Tracking (MOT) aims at tracking multiple targets simultaneously in a video data. The main application scenarios of MOT are security monitoring and autonomous driving. In these scenarios, we often need to track many targets at the same time which is not possible with only object detection or single object tracking algorithms for their lack of stability and usability. Therefore we need to explore the area of multiple object tracking. The proposed method breaks the MOT into different stages and utilizes the motion and appearance information of targets to track them in the video data. We used three different object detectors to detect the pedestrians in frames, a person re-identification model as appearance feature extractor and Kalman filter as motion predictor. Our proposed model achieves 47.6% MOT accuracy and 53.2% in IDF1 score while the results obtained by the model without person re-identification module is only 44.8% and 45.8% respectively. Our experiment results indicate the fact that a robust multiple object tracking algorithm can be achieved by splitted tasks and improved by the representative DNN based appearance features. / Detta projekt syftar till att använda djupinlärning för att lösa problemet med att följa fotgängare för autonom körning. For ligger inom datorseende och djupinlärning. Multi-Objekt-följning (MOT) syftar till att följa flera mål samtidigt i videodata. de viktigaste applikationsscenarierna för MOT är säkerhetsövervakning och autonom körning. I dessa scenarier behöver vi ofta följa många mål samtidigt, vilket inte är möjligt med endast objektdetektering eller algoritmer för enkel följning av objekt för deras bristande stabilitet och användbarhet, därför måste utforska området för multipel objektspårning. Vår metod bryter MOT i olika steg och använder rörelse- och utseendinformation för mål för att spåra dem i videodata, vi använde tre olika objektdetektorer för att upptäcka fotgängare i ramar en personidentifieringsmodell som utseendefunktionsavskiljare och Kalmanfilter som rörelsesprediktor. Vår föreslagna modell uppnår 47,6 % MOT-noggrannhet och 53,2 % i IDF1 medan resultaten som erhållits av modellen utan personåteridentifieringsmodul är endast 44,8%respektive 45,8 %. Våra experimentresultat visade att den robusta algoritmen för multipel objektspårning kan uppnås genom delade uppgifter och förbättras av de representativa DNN-baserade utseendefunktionerna.

In recent years, the field of computer vision has seen remarkable advancements, particularly with the rise of deep learning techniques. Object detection, a challenging task in image analysis, has benefited from these developments. This thesis investigates the application of object detection models, specifically You Only Look Once (YOLO), in the context of food recognition and health assessment based on the Swedish plate model. The study aims to assess the effectiveness of YOLO models in predicting the healthiness of food compositions according to the guidelines provided by the Swedish plate model. The research utilizes a custom dataset comprising 3707 images with 42 different food classes. Various preprocessing- and augmentation techniques are applied to enhance dataset quality and model robustness. The performance of the three YOLO models (YOLOv7, YOLOv8, and YOLOv9) are evaluated using precision, recall, mean Average Precision (mAP), and F1 score metrics. Results indicate that YOLOv8 showed higher performance, making it the recommended choice for further implementation in dietary assessment and health promotion initiatives. The study contributes to the understanding of how deep learning models can be leveraged for food recognition and health assessment. Overall, this thesis underscores the potential of deep learning in advancing computational approaches to dietary assessment and promoting healthier eating habits.

This study explores the application of object detection models for detecting Arctic Foxes in camera trap images, a crucial step towards automating wildlife monitoring and enhancing conservation efforts. The study involved training models on You Only Look Once version 7(YOLOv7) architecture across different locations using k-fold cross-validation technique and evaluating their performance in terms of mean Average Precision (mAP), precision, and recall. The models were tested on both validation and unseen data to assess their accuracy and generalizability. The findings revealed that while certain models performed well on validation data, their effectiveness varied when applied to unseen data, with significant differences in performance across the datasets. While one of the datasets demonstrated the highest precision (88%), and recall (94%) on validation data, another one showed superior generalizability on unseen data (precision 76%, recall 95%). The models developed in this study can aid in the efficient identification of Arctic Foxes in diverse locations. However, the study also identifies limitations related to dataset diversity and environmental variability, suggesting the need for future research to focus on training models during different seasons and having different aged Arctic Foxes. Recommendations include expanding dataset diversity, exploring advanced object detection architectures to go one step further and detect Arctic Foxes with skin diseases, and testing the models in varied field conditions.

In this thesis, we present ContextNet, a novel general object detection framework for incorporating context cues into a detection pipeline. Current deep learning methods for object detection exploit state-of-the-art image recognition networks for classifying the given region-of-interest (ROI) to predefined classes and regressing a bounding-box around it without using any information about the corresponding scene. ContextNet is based on an intuitive idea of having cues about the general scene (e.g., kitchen and library), and changes the priors about presence/absence of some object classes. We provide a general means for integrating this notion in the decision process about the given ROI by using a pretrained network on the scene recognition datasets in parallel to a pretrained network for extracting object-level features for the corresponding ROI. Using comprehensive experiments on the PASCAL VOC 2007, we demonstrate the effectiveness of our design choices, the resulting system outperforms the baseline in most object classes, and reaches 57.5 mAP (mean Average Precision) on the PASCAL VOC 2007 test set in comparison with 55.6 mAP for the baseline. / MS

Als dritthäufigstes Krebsvorkommen und zweithäufigste Krebstodesursache hat das kolorektale Karzinom (KRK) einen hohen Stellenwert für die interdisziplinäre Therapie in der Onkologie. Bei etwa 50% der Patienten befindet sich das KRK im Rektum. Die Behandlung erfolgt kurativ durch die operative Entfernung des Rektums samt der regionären Lymphknoten. Bis heute konnten keine klinischen bzw. onkologischen Vorteile der roboterassistierten Rektumresektion gegenüber der konventionell laparoskopischen Variante bewiesen werden. In dieser Arbeit wurde mithilfe maschineller Lernverfahren (Künstlicher Intelligenz, KI) ein Algorithmus trainiert, welcher bestimmte kritische anatomische Strukturen und Dissektionsebenen automatisch identifizieren kann. Damit soll zukünftig eine Assistenzfunktion etablieren werden, welche dem Chirurgen dabei helfen soll, autonome Nerven und Blutgefäße zu schonen, was das onkologische Outcome verbessern könnte. Insgesamt wurden 29 anteriore Rektumresektionen berücksichtigt, welche je in 5 OP-Phasen eingeteilt wurden (Peritoneale Inzision, Gefäßdissektion, Mediale Mobilisation, Laterale Mobilisation, Mesorektale Exzision). Etwa 500 – 2.500 Bilder wurden von jeder Phase aus den Operationsvideos extrahiert und bestimmte Strukturen wurden semantisch segmentiert. Die Leave-One-Out-Kreuzvalidierung wurde für die Algorithmus-Validierung angewendet. Als maschinelles Lernverfahren diente ein Mask R-CNN basierender Deep Learning-Algorithmus. Um die Prädiktionen evaluieren zu können, wurden die Objekterkennungs-Metriken Intersection over Union (IoU), Precision, Recall, F1 und Specificity berechnet. Gute IoU-Werte konnten bei der Instrumentenerkennung (IoU bis zu 0,82 ± 0,26), bei der Gerota’schen Faszie (IoU: 0,74 ± 0,03) und beim Mesokolon (IoU: 0,65 ± 0,05) während der medialen Mobilisation, bei der Abdominal wall (IoU: 0,78 ± 0,04) und beim Fat (IoU: 0,64 ± 0,10) während der lateralen Mobilisation und beim Peritoneum, welches beim ersten Einschnitt inzidiert wird, erreicht werden (IoU: 0,69 ± 0,22). Eine weniger präzise automatische Erkennung wurde bei der mesorektalen Faszie (IoU: 0,28 ± 0,08), beim Mesorektum (IoU: 0,45 ± 0,08), beim Kolon und Dünndarm (IoU: 0,46 ± 0,09 bzw. 0,33 ± 0,24) und der Vena mesenterica inferior (IoU: 0,25 ± 0,17) berechnet. Unzureichende Werte wurden bei den eigentlichen Dissektionslinien, den Bläschendrüsen und bei der Arteria mesenterica inferior erzielt, mit durchschnittlichen IoU-Werten kleiner 0,01 bis 0,16. Das künstliche neuronale Netzwerk erkannte zudem meist etwas ziemlich gut oder erkannte es gar nicht. Mittelgute Einzelwerte sind selten. Zusammenfassend zeigen diese Ergebnisse, dass eine KI in der Lage ist, anatomische Strukturen in laparoskopischen Aufnahmen bei einer solch komplexen OP zu erkennen. Für die KI ist es schwierig, vor allem kleinere oder hochvariabel aussehende Strukturen wie die Bläschendrüsen, Blutgefäße oder die mesorektale Faszie zu identifizieren. Es ist anzunehmen, dass die Prädiktionen mit einem größeren und diverseren Trainingsdatensatz verbessert werden können. Für Strukturen wie Dissektionslinien, für welche keine wirklichen optischen Abhebungen von anderen Strukturen im Bild bestehen, könnten andere Bereiche für die Einblendung einer Schnittführungslinie in den Bildern von Bedeutung sein. Eine zukünftige Implementierung dieser Methode in den Operationssaal im Rahmen einer Navigationsfunktion für den Chirurgen wäre demzufolge möglich.