Abstract The Automatic Identification System (AIS) is a cornerstone of modern maritime safety, enabling real-time vessel tracking and collision avoidance. However, the sheer volume of AIS data generated by global shipping presents a significant "Big Data" challenge for storage, transmission, and analysis. This paper presents a high-performance framework for the compression and visualization of large-scale vessel trajectories using Graphics Processing Unit (GPU) acceleration. By redesigning the Douglas-Peucker (DP) compression and Kernel Density Estimation (KDE) visualization algorithms for parallel execution, the researchers achieved up to a 15x speedup over traditional CPU-based methods. Experimental results on datasets comprising nearly 90 million points demonstrate that this approach maintains high visualization quality while drastically reducing computational overhead.
Introduction
In the maritime Internet of Things (IoT), AIS data serves as the primary source for monitoring vessel movements. Every large ship broadcasts its position, speed, and heading at regular intervals, creating a massive stream of trajectory data. While essential for safety and logistics, processing this data for historical analysis or real-time monitoring often hits a performance wall when using traditional serial processing.
To address these bottlenecks, this research leverages the massively parallel architecture of NVIDIA GPUs to accelerate two critical tasks: trajectory compression (to reduce data volume) and trajectory visualization (to generate density heatmaps for traffic analysis).
The Big Data Challenge in Maritime Logistics
As maritime traffic intensifies, the volume of AIS data has become overwhelming. Two primary challenges emerge:
- Algorithmic Bottlenecks: Standard algorithms like Douglas-Peucker (DP) for line simplification are inherently recursive, making them difficult to implement efficiently on parallel hardware.
- Visualization Gaps: AIS data is often discrete and "gappy." Simple point plotting fails to represent continuous movement, leading to "spotty" heatmaps that can obscure actual traffic patterns.
Proposed Methodology
The authors developed a "parallel-friendly" architecture that moves the entire data processing pipeline—from preprocessing to visualization—onto the GPU.
1. Parallel Douglas-Peucker (DP) Compression
The classic DP algorithm works by recursively splitting a line at the point of maximum distance. To adapt this for the GPU, the researchers replaced recursion with a hierarchical structure and segmented scan algorithms. This allows the GPU to calculate the "Vertical Euclidean Distance" (VED) for every point in every segment simultaneously, identifying feature points in parallel across the entire dataset.
2. GPU-Based Kernel Density Estimation (KDE)
To create smooth, continuous traffic heatmaps, the framework implements a parallel KDE algorithm.
- Linear Interpolation: To fix the "gappiness" of raw AIS data, the system performs on-the-fly linear interpolation between distant points before calculating density.
- Atomic Operations: The implementation uses
atomicAddto manage concurrent updates to the density map, preventing data hazards while maintaining high throughput.
3. Memory and Hardware Optimization
The framework optimizes data transfer by using coalesced memory access and leveraging on-chip shared memory. This minimizes the latency involved in moving data between the host CPU and the GPU device, ensuring the computational units are never starved for data.
Experimental Results and Insights
The framework was tested against three massive real-world AIS datasets from high-traffic Chinese waterways (South Channel of Yangtze River Estuary, Chengshan Jiao Promontory, and Zhoushan Islands), totaling nearly 90 million points.
- Significant Speedup: The GPU implementation achieved speedup ratios (SR) between 10.1x and 15.3x compared to serial CPU implementations. The performance advantage was most pronounced on the largest datasets.
- Compression vs. Quality: The study found that trajectories can be compressed by over 80% without noticeably degrading the quality of the resulting KDE heatmaps, provided the compression threshold is chosen appropriately.
- Kernel Selection: After comparing eight different kernel functions, the Gaussian kernel was found to produce the most natural and accurate traffic density maps.
Conclusion
This research provides a robust blueprint for modernizing maritime spatial algorithms. By moving from recursive serial logic to iterative parallel frameworks, engineers can process tens of millions of vessel data points in minutes rather than hours. Future work will focus on automating threshold selection and extending the framework to real-time anomaly detection and multi-dimensional clustering.
Summary updated