204 research outputs found
Tracking a "facer's" behavior in a public plaza
No full textThe video shows the tracking of a "facer's" behavior in a public plaza using a thermal camera (non-privacy violating) and a visualization of the tracks in a space-time cube in a 3D GIS.The tracking data is used in my PhD project on Human Movement Patterns in Smart Cities.The recording and analysis of the thermal video has been made in collaboration with Rikke Gade from the Visual Analytics of People Lab at Aalborg University.The video shows the tracking of a "facer's" behavior in a public plaza using a thermal camera (non-privacy violating) and a visualization of the tracks in a space-time cube in a 3D GIS.The tracking data is used in my PhD project on Human Movement Patterns in Smart Cities.The recording and analysis of the thermal video has been made in collaboration with Rikke Gade from the Visual Analytics of People Lab at Aalborg University
Capturing human movement patterns in public spaces
No full textNon-intrusive and non-privacy violating tracking of people by the use of thermal cameras and Computer VisionThe video shows examples of data collection of pedestrian tracks in an urban plaza using a thermal camera. The data is used in my PhD project on Human Movement Patterns in Smart Cities.The recording and analysis of the thermal videos has been made in collaboration with Rikke Gade from the Visual Analytics of People Lab at Aalborg University
Car, Bicycle and Pedestrian Detection in Adverse Weather Conditions using Lidar and Thermal Imaging
No full textIn the following work, a novel multimodal sensor data fusion on far infrared imaging and lidar is elaborated and its effects on avderse weather conditions are studied. Furthermore, a well-known problem in the thermal imaging domain is addressed, the existing large domain gap to existing datasets for the pre-training of deep neural networks. With the help of artificial thermal images the training of underrepresented object classes can be improved. The final analysis takes into account light fog, dense fog and snow during day and night. This information is finally evaluated, characterized and summarized with an outlook showing that infrared and far infrared wavelengths have strong challenges with atmospheric humidity. Both lidar and thermal imaging show degraded performance in these conditions
Detection of Plastic Pellets on Beach Sand Surface.
No full textThis project aims to contribute to the development of a robot that is capable of removing plastic pellets from beaches and coastal environments. The contribution is in the form of an investigation of possible solutions for an autonomous plastic pellet detection system and the development of the system. The system uses a detection algorithm that has been trained on a dataset consisting of RGB images of plastic pellets on beach sand surfaces. Two datasets were used, one with a simple beach environment and the other with a complex beach environment. The complex beach environment was considered complex due to the trash debris, beach material, and other artifacts that could be in that environment. The simple environment contained only sand and plastic pellets. The results from training the detection algorithm on the simple dataset showed that the detector can accurately recognize plastic pellets in a simple beach environment. The training results for the simple dataset were 0.977 for precision, 1 for recall, and 0.995 for mAP. The detection model performed less accurately on the complex dataset with a score of 0.758 for precision, 0.835 for recall, and 0.850 for mAP.This project aims to contribute to the development of a robot that is capable of removing plastic pellets from beaches and coastal environments. The contribution is in the form of an investigation of possible solutions for an autonomous plastic pellet detection system and the development of the system. The system uses a detection algorithm that has been trained on a dataset consisting of RGB images of plastic pellets on beach sand surfaces. Two datasets were used, one with a simple beach environment and the other with a complex beach environment. The complex beach environment was considered complex due to the trash debris, beach material, and other artifacts that could be in that environment. The simple environment contained only sand and plastic pellets. The results from training the detection algorithm on the simple dataset showed that the detector can accurately recognize plastic pellets in a simple beach environment. The training results for the simple dataset were 0.977 for precision, 1 for recall, and 0.995 for mAP. The detection model performed less accurately on the complex dataset with a score of 0.758 for precision, 0.835 for recall, and 0.850 for mAP
Privacy Motivated Models for Deep Learning-based Human Action Recognition in Intelligent Video Surveillance
No full textComputer-Vision Based Bend Quality Inspection for Sheet Metal Parts
No full textIn this work, the research question of how to inspect the quality of bends in 3D point cloud representations of sheet metal parts is investigated. The manufacturing system responsible for fabricating the sheet metal parts is introduced, alongside the currently available scanned data. A novel pipeline is developed, which automatically pre-processes the scanned point clouds, triangulates the positions of the bends and measures their total inclination angle, as well as their arc length. The method is tested on ideal data, which proved that it is conceptually sound. Subsequent experiments are conducted to test the method's robustness to noise, as well as its ability to provide realistic results on the real scans
Correction to:Trends in use of attention deficit hyperactivity disorder medication among children and adolescents in Scandinavia in 2010–2020 (European Child & Adolescent Psychiatry, (2023), 32, 10, (2049-2056), 10.1007/s00787-022-02034-2)
No full textIn the original publication, the second author name published incorrectly. The correct name is Rikke Wesselhoeft. The original article has been corrected.</p
Where can I go? Deep multi-modal scene understanding for outdoor navigation
No full textThis project delves into deep learning-based computer vision for scene understanding in the context of autonomous outdoor navigation. Rather than relying on specific scene-dependent semantic categories, we take an affordance-based approach, proposing to parse egocentric images in terms of how a vehicle or robot can drive in them. We use a SegNet-based image segmentation network as our building block for classifying pixels into 3 driveability levels, and explore soft labelling, pixel-wise loss weighting, and deep adaptive fusion schemes to penalize severe mistakes during learning, improve segmentation in regions of interest, and incorporate infrared and depth data into the prediction. The proposed training schemes and multi-modal architecture are evaluated on 9 public datasets, showing promising results across unstructured forested environments, urban driving scenes, and multi-view hand-held captures
Sports Tracking by Correlating Thermal Video with Inertia Data
No full textThe aim of this project is to correlate thermal video with an accelerometer and be able to connect tracklets made from a video tracker to trajectories for each person.The project resulted in developing a pipeline capable of detecting players, tracking them, calculate acceleration in the video to match with the acceleration from the accelerometers and from this be able to match accelerometer to tracklets.The report covers all the blocks in this process: Detection, tracking and matching, where each block have been designed, implemented and tested.The proof-of-concept system has been tested on a dataset made of half a field in Gigantium sports arena, where the participants played football. The system shows promising results and that it is possible to correlate thermal video with an accelerometer, but the tracking system requires more work to know if it can be implemented in an automatic system
Synthesizing Hard Training Data From Latent Hierarchical Representations
Get PDFThis paper introduces a framework for creating augmented hard samples, which are new images created to resemble those that a classifier will struggle to classify. This is used for data from an automatic visual defect inspection system, specifically images of vials with and without chipped glass. The hard samples were found by training ConvNeXt classifiers and using the confidences of the classifiers on the training dataset. VQ-VAE2 was used to obtain the latent representations of the hard samples, and a PixelSnail model was used to create new high-frequency details while retaining low-frequency details. The PixelSnail model was pre-trained on a large amount of non-defect images. The augmentation method was applied to a dataset of 200 images and was evaluated by training classifiers to test the effect of using the augmented hard samples. Introducing the augmented hard samples into the dataset was found to improve classifier performance, measured in Area Under Curve (AUC) of the Receiver Operator Characteristic (ROC) curve, from 0.953 to 0.973. The method was tested with augmenting both defect and non-defect images, and the somewhat surprising conclusion is that while using augmented defect images didn’t yield improvements, augmenting non-defect images did.This paper introduces a framework for creating augmented hard samples, which are new images created to resemble those that a classifier will struggle to classify. This is used for data from an automatic visual defect inspection system, specifically images of vials with and without chipped glass. The hard samples were found by training ConvNeXt classifiers and using the confidences of the classifiers on the training dataset. VQ-VAE2 was used to obtain the latent representations of the hard samples, and a PixelSnail model was used to create new high-frequency details while retaining low-frequency details. The PixelSnail model was pre-trained on a large amount of non-defect images. The augmentation method was applied to a dataset of 200 images and was evaluated by training classifiers to test the effect of using the augmented hard samples. Introducing the augmented hard samples into the dataset was found to improve classifier performance, measured in Area Under Curve (AUC) of the Receiver Operator Characteristic (ROC) curve, from 0.953 to 0.973. The method was tested with augmenting both defect and non-defect images, and the somewhat surprising conclusion is that while using augmented defect images didn’t yield improvements, augmenting non-defect images did
- …
