Automatic video quality assessment has many potential use cases in today’s video-filledsociety, for example, when trying to find highlights in a video. This thesis studies the possibilityof extracting the best segments from a video automatically based on five selected metrics:sharpness, colorfulness, contrast, stability, and aesthetics. Multiple different methods from eachmetric category were compared against each other using datasets with subjective ratings ofimages from KADID-10k and videos from LIVE-Qualcomm. The best method from eachcategory was combined into a single video quality score. The combination was done through aweighted sum, obtained from a least-square fit on the subjective scores of a training dataset(KonViD-1k). The segment of a video with the highest average quality score was chosen as thehighlight. The video quality score achieved Spearman correlations of 0.67 and 0.7 whenevaluated on two validation datasets (KonViD-150k-B and LIVE-VQC). In conclusion, themetrics work well, but are currently too slow for the intended target platform (mobile devices)and thus future work should focus on improving their performance.

There is a renewed fücus on active modes of transportation given their multiple advantages, whether für human health or the environment in general. Interest has grown especially in 2020 after the COVID-19 pandemic, when several cities quicldy implemented temporary facilities for walking and cycling in the context of physical distancing. Several measures piggyback.ed on existing programs such as the Montreal initiative for complete streets ('rues conviviales' or 'social/festive streets'') that selects streets each year für pilot projects and a final design implementation over a three-year period This resulted in seasonal pedestrianization of about ten streets each year since 2020. Though active transportation brings together pedestrians and cyclists und.er a large umbrella, these users have very different characteristics and tbere may be conflicts of use if mixed in tbe same space. Cycling is thus generally forbidden on pedestrian streets. Despite these rules, there is cycling traffic on pedestrian streets as cyclists also enjoy car-free facilities, especially when pedestrian traffic is low, which generates complaints by pedestrians. To reconcile and help botb categories of users coexist, two Montreal boroughs tried a new rule in the Summer of 2021, to 1et cyclists bik.e at walldng speed on pedestrian streets while avoiding conflicts with pedestrians. There are few studies on cyclist-pedestrian interactions, and, to the best ofthe authors' knowledge, none on interactions in pedestrian streets. This work aims to study the coexistence or cohabitation of pedestrians and cyclists in several pedestrian streets through video-based analysis. Data were collected at several sites and on several days during the Summer of 2021 along three different pedestrian streets, two of them. allowing cycling, to assess how cyclists and pedestrians interact, whether cycling is allowed or not.

This thesis focuses on the development and utilization of a virtual reality bicycle simulator for the purposes of traffic safety research. The bicycle simulator is developed using a virtual reality head mounted display (HMD) and commercially available hardware and software in the Unity framework. An experiment was conducted by exposing 24 participants to select scenarios in a virtual environment that were constructed to imitate a real-world location in Lund, Sweden. Observing the effects in term of immersion and cybersickness in relation to development. Recordings from drones were used to capture real life traffic from the location that were tracked and implemented into Unity for increased realism. The participants answered questionnaires incorporating VRSQ: Virtual reality sickness questionnaire and SPES: The spatial presence experience scale. The results indicate that the induced cybersickness is similar to, but slightly lower than that of average mean simulator sickness questionnaire (SSQ) results. SPES showed generally positive results, average answer being 3.6 on a scale from 1-5.

Ice hockey has been identified as a sport with a high risk for concussions due to it being highly physical. Improvements have been made over the years to improve the protective gear for the players effectively eliminating more severe traumatic brain injuries (TBIs) however mild concussions (mTBI) are still prevalent to this day. One way of predicting these injuries in the recent years is the usage of finite element (FE) analysis to recreate impacts to study the effects of said head kinematics and strain parameters using validated FE models of the brain. In this thesis, video analysis was done on five cases resulting in a concussion and five cases not resulting in concussion from both the Swedish Hockey League (SHL) and the National Hockey League (NHL) to extract the initial velocities and positioning. The average velocity for the injured player and attacking player was 5.12 m/s and 5.08 m/s respectively for the impacts resulting in a concussion. Additionally, the average velocity for the injured player and attacking player was 4.80 m/s and 5.51 respectively for the impacts not resulting in a concussion. The video analysis methodology was also validated using a dataset from a football game and resulted in an average error of 25.4%. The impacts were recreated using the extracted velocities and initial positions in LS-PrePost with the full body FE model THUMS v.4.02 representing a 50th percentile adult male fitted with a helmet previously developed by master thesis students at KTH. The simulations were ran using LS-DYNA. The head kinematics from the head’s center of gravity and brain strain measured called Maximum Principal Strain (MPS) were extracted. Using the MPS values, the 95th percentile was calculated to then determine the likelihood of concussion. The MPS95 ranged from 0.18 to 0.63 for the cases resulting in a likelihood of concussion of 7.48 to 100%. For the cases not resulting in a concussion, the MPS95 values ranged from 0.20 to 0.45 resulting in likelihood of concussion 11.1 to 80.5%. The head kinematics extracted reported similar outcome in terms of risk of suffering concussions. The varying results can be pointed to the drawbacks in the methodology such as the error of the video analysis and the positioning of FE models. / Ishockey har fastställts som en sport med höga risker för hjärnskakning på grund av dess fysiska natur. Med åren har förbättringar gjorts på skyddsutrustningen vilket har avlägsnat allvarligare hjärnskador (TBI) från sporten, men dessvärre förekommer mildare hjärnskakningar än idag. På senare år har användning av finita element (FE) metoder använts för att återskapa kollisionerna för att förutspå dessa skador. Detta görs genom att använda validerade FE modeller av hjärnan för att analysera kinematiken samt töjningarna. I detta arbete gjordes en videoanalys av fem fall som resulterade i hjärnskakning samt fem fall där kollisionen inte resulterade i en hjärnskakning från både den Svenska Hockey Ligan (SHL) och "National Hockey League" (NHL). Från videoanalysen togs de initiala hastigheterna samt positionerna fram. Den genomsnittliga hastigheten för den skadade spelaren respektive attackerande spelaren var 5.12 m/s och 5.02 m/s för fallen där hjärnskakning hade skett. För fallen där hjärnskakning inte hade skett var hastigheten för den skadade respektive attackerande spelaren var 4.80 m/s samt 5.51 m/s. Videoanalysmetoden validerades genom att utföra videoanalysis på en databas från en fotbollsmatch. Valideringen resulterade i ett genomsnittligt fel på 25.4%. Kollisionerna återskapades genom att använda de extraherade initiala hastigheterna och positionerna i LS-PrePost med FE helkroppsdockan THUMS v.4.02 som representerade en 50:e percentil vuxen man utrustad med en ishockeyhjälm som utvecklats av tidigare masterexamensstudenter på KTH. Simuleringarna gjordes genom LS-DYNA. Huvudkinematiken från huvudets tyngdpunkt samt de maximala principiella töjningarna (MPS) i hjärnan extraherades. Från de extraherade MPS värdena kunde den 95:e percentilen uträknas för att bestämma sannolikheten att hjärnskakning sker. För fallen där hjärnskakning skett, varierade MPS95 värdet mellan 0.18 till 0.63 vilket motsvarar en risk för hjärnskakning på 7.48% respektive 100%. För fallen där hjärnskakning inte hade skett varierade MPS95 värdet från 0.20 till 0.45 vilket motsvarar 11.1% respektive 80% risk för hjärnskakning. Den extraherade huvudkinematikens risk för hjärnskakning överensstämde för det mesta med MPS95 sannolikheterna. De varierande resultaten kan hänvisas till de brister i metoden som exempelvis felet i videoanalysen samt positioneringen av FE modeller.

<p dir="ltr">his research investigates the detection of pre-crime events, specifically targeting behaviors indicative of shoplifting, through the advanced analysis of CCTV video data. The study introduces an innovative approach that leverages augmented human pose and emotion information within individual frames, combined with the extraction of activity information across subsequent frames, to enhance the identification of potential shoplifting actions before they occur. Utilizing a diverse set of models including 3D Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), Recurrent Neural Networks (RNNs), and a specially developed transformer architecture, the research systematically explores the impact of integrating additional contextual information into video analysis.</p><p dir="ltr">By augmenting frame-level video data with detailed pose and emotion insights, and focusing on the temporal dynamics between frames, our methodology aims to capture the nuanced behavioral patterns that precede shoplifting events. The comprehensive experimental evaluation of our models across different configurations reveals a significant improvement in the accuracy of pre-crime detection. The findings underscore the crucial role of combining visual features with augmented data and the importance of analyzing activity patterns over time for a deeper understanding of pre-shoplifting behaviors.</p><p dir="ltr">The study’s contributions are multifaceted, including a detailed examination of pre-crime frames, strategic augmentation of video data with added contextual information, the creation of a novel transformer architecture customized for pre-crime analysis, and an extensive evaluation of various computational models to improve predictive accuracy.</p>

Die Datenbestände lokaler Fernsehsender umfassen oftmals mehrere zehntausend Videokassetten. Moderne Verfahren werden benötigt, um derartige Datenkollektionen inhaltlich automatisiert zu erschließen. Das Auffinden relevanter Objekte spielt dabei eine übergeordnete Rolle, wobei gesteigerte Anforderungen wie niedrige Fehler- und hohe Detektionsraten notwendig sind, um eine Korruption des Suchindex zu verhindern und erfolgreiche Recherchen zu ermöglichen. Zugleich müssen genügend Objekte indiziert werden, um Aussagen über den tatsächlichen Inhalt zu treffen. Diese Arbeit befasst sich mit der Anpassung und Optimierung bestehender Detektionsverfahren. Dazu wird ein auf die hohen Leistungsbedürfnisse der Videoanalyse zugeschnittenes holistisches Workflow- und Prozesssystem mit der Zielstellung implementiert, die Entwicklung von Bilderkennungsalgorithmen, die Visualisierung von Zwischenschritten sowie deren Evaluation zu ermöglichen. Im Fokus stehen Verfahren zur strukturellen Zerlegung von Videomaterialien und zur inhaltlichen Analyse im Bereich der Gesichtsdetektion und Fußgängererkennung. / The data collections of local television stations often consist of multiples of ten thousand video tapes. Modern methods are needed to exploit the content of such archives. While the retrieval of objects plays a fundamental role, essential requirements incorporate low false and high detection rates in order to prevent the corruption of the search index. However, a sufficient number of objects need to be found to make assumptions about the content explored. This work focuses on the adjustment and optimization of existing detection techniques. Therefor, the author develops a holistic framework that directly reflects on the high demands of video analysis with the aim to facilitate the development of image processing algorithms, the visualization of intermediate results, and their evaluation and optimization. The effectiveness of the system is demonstrated on the structural decomposition of video footage and on content-based detection of faces and pedestrians.

La douleur est une expérience multidimensionnelle comportant des aspects sensoriels, émotionnels et cognitifs. Théoriquement, des méthodes de mesures comportementales, physiologiques, neurophysiologiques et sensorielles peuvent quantifier la douleur. Peu d’études ont étudié la validation des mesures utilisées en médecine vétérinaire. La recherche combine les travaux de Maîtrise et de Doctorat, traite en partie de la validité de méthodes. Dans cet objectif, nos travaux de recherche étudiaient la validité de méthodes comportementales, physiologiques et neurophysiologiques usuelles pour la mesure de la douleur en comparant les expressions de douleur (vache et chien) chez des animaux contrôle par comparaison à des animaux sous analgésie préventive ou sous traitement curatif suivant une douleur induite par chirurgie (modèles de douleur viscérale bovine ou orthopédique canine) ou causée par une maladie naturelle (arthrose canine). Une première étude comparait les mesures de la douleur entre les vaches du groupe placebo et celles sous analgésie postopératoire sur une durée de 21 jours suivant l’induction d’une douleur viscérale chronique. Les vaches du groupe placebo ont présenté une plus forte sensibilité à la douleur et une diminution de la noradrénaline et de la transthyrétine mesurées dans le liquide cérébro-spinal, une diminution de l’activité motrice (AM) (moindre que dans les groupes avec analgésie), de l’agitation enregistrée par vidéo-analyse et une augmentation du stress selon la mesure de l’activité électrodermique (AED). Les méthodes d’intérêt identifiées étaient les marqueurs spinaux, la mesure de la sensibilisation, de comportements par vidéo-analyse et de l’AM par bio-télémétrie. En utilisant des méthodes semblables à celles précédemment décrites, deux études expérimentales de douleur orthopédique ont été réalisées afin de comparer les réponses à la douleur entre des chiens traités avec une analgésie préventive (opioïdes et anti-inflammatoires, étude #2) ou un biphosphonate (tiludronate, étude #3) par comparaison à des chiens contrôles. Seules les échelles de douleur étaient différentes entre les études de recherche. Pour l’étude #2, les ii chiens sous analgésie ont présenté de plus faibles scores de douleur mesurés avec l’échelle de douleur nommée 4A-VET et ceci simultanément à une faible réponse de l’AED une heure après la chirurgie de trochléoplastie. La fréquence du comportement spontané de ‘la marche avec plein appui de la patte opérée’ mesurée à l’aide de la vidéo-analyse augmentait chez les chiens sous analgésie préventive 24 heures après la chirurgie. L’étude #3 démontrait surtout l’apparition de sensibilisation centrale (à la fois par l’évaluation sensorielle quantitative et les marqueurs spinaux) chez les chiens contrôle, 56 jours après l’induction d’arthrose chirurgicale. Ainsi, les chiens traités avec le tiludronate ont présenté une différence sur la substance P et la transthyrétine cérébro-spinale, une diminution de la sensibilisation périphérique, plus d’appui de la patte opérée lors de la marche selon la mesure du pic de force verticale (PFV), une augmentation de la fréquence de ‘la marche avec plein appui de la patte opérée’. La sensibilisation centrale était associée à la diminution de PFV, et une augmentation de l’AED et du comportement spontané de ‘la marche avec plein appui de la patte opérée’. Pour l’étude #4, la validité et la sensibilité des méthodes ont été évaluées dans une condition d’arthrose naturelle chez des chiens traités avec une diète enrichie en moule verte, un produit ayant des effets anti-inflammatoires et chondroprotecteurs attendus. Les chiens traités présentaient une diminution des scores de douleur via l’échelle nommée CSOM, une augmentation de PFV et une augmentation de l’AM. Dans l’ensemble, les résultats confirment que la vidéo-analyse évaluait la douleur de façon objective et pour des modèles différents de douleur et les marqueurs spinaux sont prometteurs. Le PFV était spécifique de la douleur orthopédique. La sensibilisation était présente lors de douleur pathologique. L’AED n’est pas valide pour la mesure de la douleur. La baisse d’AM suggèrerait un comportement de douleur. Les études étaient exploratoires pour les échelles de douleur en raison de leur niveau (débutant) de développement et du manque d’informations sur les qualités métrologiques de ces mesures. / Pain is a multidimensional experience involving sensitive, emotional and cognitive components. Theoretically, there are multiple methods by which pain can be assessed including sensitive, behavioural, physiological, or neurophysiological measurements. However, little work has been done to validate these measurements in veterinary medicine. The presented research program including both Master and Doctorate works was intended to address partially this paucity of research. For this purpose, our work would validate some behavioural and physiological methods of pain assessment by contrasting pain expressions (cows and dogs) in painful animals (negative control) and animals treated with preventive analgesic or curative treatment following surgery-induced (bovine visceral and canine orthopaedic models) pain or natural occurring disease (osteoarthritis in dog). A pain study was first conducted to compare measurements of placebo treated-cows with postoperative analgesic treated-cows during 21 days following surgical induction of sustained visceral pain. Placebo treated-cows were found to have increased pain sensitization and decreased concentration of cerebrospinal fluid noradrenaline and transthyretin, less motor activity (but higher than in analgesic groups), more restlessness recorded with video-analysis and increased partially stress with measurement of electrodermal activity (EDA). This first study allowed a selection of methods of interest for pain evaluation including spinal biomarkers, measurement of sensitization, behavioural recording with video-analysis and motor activity with biotelemetry. Therefore, two canine pain experiments, with use of similar methods of pain assessment presented above, were performed to compare responses to pain between preventive analgesics treated-dogs (opioids and anti-inflammatory drug, study #2) or a bisphosphonate (tiludronate in study #3) with placebo-treated dogs. Only the pain scales were different among the projects. For project #2, analgesic treated-dogs were found to have lower pain scores measured with the so-called 4A-VET postoperative pain scale while simultaneously exhibiting reduction of EDA response up to 1 hour following trochleoplasty. In addition, the occurrence rate of the spontaneous behaviour ‘Walking with full weight bearing of the operated leg’ recorded with video-analysis, was higher in analgesic treated-dogs when compared with the placebotreated dogs at 24 hours post trochleoplasty. The pain study #3 was then conducted and demonstrated central sensitization (assessed with quantitative sensory testing and spinal biomarkers) in all control dogs at 56 days post induction of the canine osteoarthritis pain model. Nevertheless, tiludronate treated-dogs were found to have different spinal biomarkers (substance P and transthyretin), decreased peripheral sensitization, more peak vertical force (PVF), which is a kinetic gait parameter, and increased occurrence rate of ‘Walking with full weight bearing of the operated leg’. Interestingly, the central sensitization was associated negatively with PVF and positively with both EDA and ‘Walking with full weight bearing of the operated leg’. Finally, a fourth pain study was conducted to examine whether some of the methods performed validity and sensitivity in clinical condition with osteoarthritic dogs. For this purpose, osteoarthritic dogs were treated with a green-lipped mussel enriched-diet, having both anti-inflammatory and chondroprotective expected activities. The treated-dogs were found to have low pain scores measured with the pain scale for owner named CSOM, increased PVF and motor activity. Indeed, CSOM scores were associated with both PVF and motor activity. Taken together, the results suggest that video-analysis would assess pain expression through objective, predictive and unique evaluation whatever the species or the model, whereas spinal biomarkers are promising. The PVF changes were related to orthopaedic pain. Sensitization appeared to be common to the pathological pain pattern. The EDA was not validated for pain assessment in animals. Decreased motor activity is pain suggestive. Psychometric evaluation of the pain scales remained only exploratory at this (early) stage of development and knowledge of the present pain scales.

This thesis describes the development of methods for automatically obtaining coarse gaze direction estimates for pedestrians in surveillance video. Gaze direction estimates are beneficial in the context of surveillance as an indicator of an individual's intentions and their interest in their surroundings and other people. The overall task is broken down into two problems. The first is that of tracking large numbers of pedestrians in low resolution video, which is required to identify the head regions within video frames. The second problem is to process the extracted head regions and estimate the direction in which the person is facing as a coarse estimate of their gaze direction. The first approach for head tracking combines image measurements from HOG head detections and KLT corner tracking using a Kalman filter, and can track the heads of many pedestrians simultaneously to output head regions with pixel-level accuracy. The second approach uses Markov-Chain Monte-Carlo Data Association (MCMCDA) within a temporal sliding window to provide similarly accurate head regions, but with improved speed and robustness. The improved system accurately tracks the heads of twenty pedestrians in 1920x1080 video in real-time and can track through total occlusions for short time periods. The approaches for gaze direction estimation all make use of randomised decision tree classifiers. The first develops classifiers for low resolution head images that are invariant to hair and skin colours using branch decisions based on abstract labels rather than direct image measurements. The second approach addresses higher resolution images using HOG descriptors and novel Colour Triplet Comparison (CTC) based branches. The final approach infers custom appearance models for individual scenes using weakly supervised learning over large datasets of approximately 500,000 images. A Conditional Random Field (CRF) models interactions between appearance information and walking directions to estimate gaze directions for head image sequences.

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal.

Honigbienen (Apis mellifera L.) sind in der Lage mit ihren Komplexaugen visuelle Muster wahrzunehmen und die Musterinformation im Zentralen Nervensystem zu speichern und für Ähnlichkeitsbewertungen wieder abzurufen. Die vorliegende Arbeit zeigt klare Evidenz gegen eine ausschließliche Bewertung von Schwarzweißmustern mit Hilfe von Template-Matching-Mechanismen. Mit systematisch abgewandelten Dressurparadigmen trainierte Bienen bewerteten Muster unabhängig von der erfolgten Dressur stets bevorzugt gemäß eher grober Mustereigenschaften, wie zum Beispiel die Parameter "schwarzer Musterzentralbereich" und "Musterzerstreutheit". Veränderte man in einem weiteren Versuchansatz die Musterinformation der Schwarzweißmuster zudem gezielt durch geeignete Fourier-Filterung, zeigte sich, dass Bienen zur Musterdiskriminierung bereits die Frequenzinformation von 2 - 8 Schwingungen/Bildbreite genügte. Diese Unschärfe der bewerteten Bildinformation ließ sich nicht ausschließlich aus den optischen Eigenschaften des visuellen Apparates der Bienen ableiten. Videodokumentationen und Einzelbildanalyse des Flugverhaltens der Bienen vor den Mustern ergaben zudem keinerlei Hinweise für eine Nutzung des Flugverhaltens als Bewertungsgrundlage zur Musterdiskriminierung. Die erhaltenen Ergebnisse zur Musterdiskriminierung wurden vor dem Hintergrund eines ökonomischen Entscheidungsmodells für menschliches Verhalten, den Frugalheuristiken, diskutiert und Hinweise auf eine ökonomische Bewertungsstrategie der Bienen entsprechend einer Take-The-Best-Heuristik gefunden. / Honeybees (Apis mellifera L.) are able to perceive visual patterns through their compound eyes and store the visual information in the central nervous system for subsequent use in pattern discrimination tasks. This thesis provides clear evidence against the assumption that pattern discrimination relies exclusively on template matching mechanisms. Bees discriminated pairs of patterns preferential using extracted pattern parameters. Within this thesis the preferred parameters of the bees following the training paradigms were coarse parameters such as "black centre" and "pattern disruption". In experiments with Fourier filtered patterns the frequency information of the patterns were additionally reduced. The results showed that bees could discriminate patterns using only 2 - 8 cycles/pattern-width of the frequency information. The fuzziness of the exploited visual information could not be assigned to restrictions of the visual system of bees. Additional documentation and single picture analysis of the videotaped flight behaviour in front of the patterns provided no evidence for bees using their flight behaviour in order to enhance the pattern discrimination ability. Application of economic human decision models (frugal heuristics) to the behavioural results showed clues that bees'' decisions could be explained with the help of the Take-The-Best-heuristic.