Files
Abstract
This dissertation presents the systematic design and development of datasets, algorithms, and an AI pipeline specifically curated for real-time trajectory prediction and anomaly detection in highway environments. These innovations are meticulously optimized for embedded-edge systems, ensuring timely outputs for safety and surveillance applications. First, the dissertation presents DeepTrack: Lightweight Deep Learning for Vehicle Path Prediction in Highways, a deep learning model tailored for edge systems. This model uniquely employs Agile Temporal Convolutional Networks (ATCNs) rather than the traditionally-used Long Short-Term Memory (LSTM) networks to encapsulate vehicle dynamics. Not only does DeepTrack boast equivalent or superior accuracy to leading trajectory prediction models, but it also shines in its diminished model size and reduced computational intensity, making it ideal for embedded edge systems. Distinctly, vehicle interactions are interpreted through ATCNs instead the frequently-associated LSTM in time series analysis. A hallmark of ATCN is its depthwise convolution, which significantly curtails model complexity in comparison to LSTMs, both in size and operational demands. Experimental results show that DeepTrack cuts Average Displacement Error (ADE) by 12.23%, reduces the Final Displacement Error (FDE) by 2.69%, and also decreases both the number of operations and the model size by approximately 21.67% and 43.13%, respectively, compared to then State of the Art (SotA) trajectory prediction algorithm.Subsequently, Carolina Highway Dataset (CHD), a unique highway trajectory dataset captured from two distinct Points Of View (POVs) – high-angle and eye-level is introduced. While numerous vehicle trajectory datasets exist, most lack the diversity of driving scenes that capture various highway designs, merging lanes, and configurations. CHD, however, stands out by offering data from 1.6 million frames and 338,000 vehicle trajectories recorded in highway videos, encapsulating both eye-level and high-angle perspectives from eight strategically selected locations across the Carolinas. These locations, accompanied by meticulous timing and camera angles, ensure a comprehensive representation of road geometries, traffic trends, lighting variations, and diverse driving behaviors. Additionally, PishguVe – a SotA vehicle trajectory prediction architecture that uses attention-based graph isomorphism and convolutional neural networks is presented next. When tested, PishguVe excelled by outperforming pre-existing SotA algorithms across bird's-eye, eye-level, and high-angle POV trajectory datasets. Notably, on the NGSIM dataset, it achieved a commendable improvement of 12.50% in Average Displacement Error (ADE) and 10.20% in Final Displacement Error (FDE) compared to the current SotA. Against the top-performing models on CHD, PishguVe demonstrated superior results, reducing the ADE and FDE on eye-level as well as high-angle data.The final contribution details VegaEdge: A Confluence AI Approach for Video Anomaly Detection at the Edge in Real-Time Highway Safety. It commences with the introduction of the Carolinas Anomaly Dataset (CAD), aiming to fill the prevalent void in datasets specifically designed for highway anomalies. The significance of vehicle anomaly detection is underscored, emphasizing its pivotal role in various highway safety applications, including accident prevention, rapid response, traffic flow optimization, and work zone safety. In alignment with this vision, a novel lightweight technique for vehicle anomaly detection is put forth, harnessing the prowess of trajectory prediction. The proposed methodology adeptly spots vehicles that deviate from anticipated paths, pinpointing potential highway risks across diverse camera perspectives, and utilizing real-world highway datasets.Further, the edge detection framework, VegaEdge is introduced. It represents a refined AI confluence that strategically chooses agile algorithms and techniques. This selection is adept at tasks ranging from detection and tracking to trajectory prediction and anomaly identification in real-time, catering to contemporary security and surveillance needs in modern highways via edge-focused IoT-embedded platforms. In addition, VegaEdge, when deployed on an embedded IoT platform, efficiently processes 738 trajectories per second in standard highway environments, illustrating its adaptability and efficiency.