Driving Behavior Analysis (DBA) plays a pivotal role in designing intelligent transportation systems, enhancing road safety, and advancing Autonomous Vehicles (AVs). Driver identification, as a key aspect of DBA, has the potential to provide unprecedented opportunities for enhanced security and driver profiling. However, the current solutions for driver identification suffer from demanding extensive data collection, limited scalability, and inadequate generalization. Furthermore, DBA is also essential for training AVs, addressing the main challenges they face: accurately perceiving their surroundings to make informed decisions and to navigate safely, and effectively handling unforeseen scenarios. In the first part of this thesis, we concentrate on behavior analysis for driver identification and verification and design two novel schemes aiming to reduce data dependency and enhance the generalization ability of existing approaches. First, we propose a novel driver identification model, called DriverRep, which reduces data dependency by presenting a fully unsupervised triplet loss training. DriverRep is the first model that extracts the latent representations associated with each driver, called driver embeddings, in an unsupervised manner. In addition, we develop a novel model to tackle driver verification and impostor detection tasks based on DBA and extracted driver embeddings. In the second part, we focus on behavior prediction for AVs and their surrounding agents. First, we tackle behavior prediction in dynamic and complex scenarios by introducing three novel prediction models for forecasting drivers intentions and behaviors at unsignalized intersections. We then address social reasoning by proposing a novel prediction model that analyzes agent interactions using graph neural networks, making the scene understanding process more informative for AVs. Our proposed prediction model, called STAG, explicitly activates social modeling with a directed graph representation while considering spatial and temporal inter-agent correlations. We further design a novel prediction system, namely CAPHA, which conditions the future behavior of agents on grid-based plans modeled as a Markov decision process and solves the prediction task via inverse reinforcement learning to produce scene compliant behaviors. Moreover, we introduce a novel goal-based prediction model, called GMP, which encodes interactions between agents and dynamic and static context information to estimate the distribution of target goals, efficiently considering the inherent uncertainty in agents behavior. Extensive quantitative and qualitative comparisons have been conducted between the developed solutions and related benchmark schemes using various scenarios and environments. The obtained results demonstrate the potential of these solutions for the understudy tasks of DBA and real-world applications.

Object-centric geometric perception aims at extracting the geometric attributes of 3D objects. These attributes include shape, pose, and motion of the target objects, which enable fine-grained object-level understanding for various tasks in graphics, computer vision, and robotics. With the growth of 3D geometry data and 3D deep learning methods, it becomes more and more likely to achieve such tasks directly using 3D input data. Among different 3D representations, a 3D point cloud is a simple, common, and memory-efficient representation that could be directly retrieved from multi-view images, depth scans, or LiDAR range images. Different challenges exist in achieving object-centric geometric perception, such as achieving a fine-grained geometric understanding of common articulated objects with multiple rigid parts, learning disentangled shape and pose representations with fewer labels, or tackling dynamic and sequential geometric input in an end-to-end fashion. Here we identify and solve these challenges from a 3D deep learning perspective by designing effective and generalizable 3D representations, architectures, and pipelines. We propose the first deep pose estimation for common articulated objects by designing a novel hierarchical invariant representation. To push the boundary of 6D pose estimation for common rigid objects, a simple yet effective self-supervised framework is designed to handle unlabeled partial segmented scans. We further contribute a novel 4D convolutional neural network called PointMotionNet to learn spatio-temporal features for 3D point cloud sequences. All these works advance the domain of object-centric geometric perception from a unique 3D deep learning perspective. / Doctor of Philosophy / 3D sensors these days are widely equipped on various mobile devices like a depth camera on iPhone, or laser LiDAR sensors on an autonomous driving vehicle. These 3D sensing techniques could help us get accurate measurements of the 3D world. For the field of machine intel- ligence, we also want to build intelligent system and algorithm to learn useful information and understand the 3D world better. We human beings have the incredible ability to sense and understand this 3D world through our visual or tactile system. For example, humans could infer the geometry structure and arrangement of furniture in a room without seeing the full room, we are able to track an 3D object no matter how its appearance, shape and scale changes, we could also predict the future motion of multiple objects based on sequential observation and complex reasoning. Here my work designs various frameworks to learn such 3D information from geometric data represented by a lot of 3D points, which achieves fine-grained geometric understanding of individual objects, and we can help machine tell the target objects' geometry, states, and dynamics. The work in this dissertation serves as building blocks towards a better understanding of this dynamic world.

No description available.

This thesis focuses on addressing the limitations of existing human motion prediction models by extending the prediction horizon to very long-term forecasts. The objective is to develop a model that achieves one of the best stable prediction horizons in the field, providing accurate predictions without significant error increase over time. Through the utilization of quantization based models our research successfully achieves the desired objective with the proposed aligned version of Mean Per Joint Position Error. The first of the two proposed models, an attention-based Vector Quantized Variational AutoEncoder, demonstrates good performance in predicting beyond conventional time boundaries, maintaining low error rates as the prediction horizon extends. While slight discrepancies in joint positions are observed, the model effectively captures the underlying patterns and dynamics of human motion, which remains highly applicable in real-world scenarios. Furthermore, our investigation into a correlation-based Vector Quantized Variational AutoEncoder, as an alternative to attention-based one, highlights the challenges in capturing complex relationships and meaningful patterns within the data. The correlation-based VQ-VAE’s tendency to predict flat outputs emphasizes the need for further exploration and innovative approaches to improve its performance. Overall, this thesis contributes to the field of human motion prediction by extending the prediction horizon and providing insights into model performance and limitations. The developed model introduces a novel option to consider when contemplating long-term prediction applications across various domains and sets the foundation for future research to enhance performance in long-term scenarios. / Denna avhandling fokuserar på att hantera begränsningarna i befintliga modeller för förutsägelse av mänskliga rörelser genom att utöka förutsägelsehorisonten till mycket långsiktiga prognoser. Målet är att utveckla en modell som uppnår en av de bästa stabila prognoshorisonterna inom området, vilket ger korrekta prognoser utan betydande felökning över tiden. Genom att använda kvantiseringsbaserade modeller uppnår vår forskning framgångsrikt det önskade målet med den föreslagna anpassade versionen av Mean Per Joint Position Error. Den första av de två föreslagna modellerna, en uppmärksamhetsbaserad Vector Quantized Variational AutoEncoder, visar goda resultat när det gäller att förutsäga bortom konventionella tidsgränser och bibehåller låga felfrekvenser när förutsägelsehorisonten förlängs. Även om små avvikelser i ledpositioner observeras, fångar modellen effektivt de underliggande mönstren och dynamiken i mänsklig rörelse, vilket förblir mycket tillämpligt i verkliga scenarier. Vår undersökning av en korrelationsbaserad Vector Quantized Variational AutoEncoder, som ett alternativ till en uppmärksamhetsbaserad sådan, belyser dessutom utmaningarna med att fånga komplexa relationer och meningsfulla mönster i data. Den korrelationsbaserade VQ-VAE:s tendens att förutsäga platta utdata understryker behovet av ytterligare utforskning och innovativa metoder för att förbättra dess prestanda. Sammantaget bidrar denna avhandling till området för förutsägelse av mänskliga rörelser genom att utöka förutsägelsehorisonten och ge insikter om modellens prestanda och begränsningar. Den utvecklade modellen introducerar ett nytt alternativ att ta hänsyn till när man överväger långsiktiga prediktionstillämpningar inom olika områden och lägger grunden för framtida forskning för att förbättra prestanda i långsiktiga scenarier.