TL;DR
This paper introduces a novel computational architecture, GC-MTT, that synergistically combines Cloud computing and Graphics Processing Units (GPUs) to achieve real-time detection and tracking of multiple moving targets in Wide Area Motion Imagery (WAMI). This approach significantly accelerates processing compared to traditional CPU-based methods, making large-scale surveillance data more manageable and actionable.
What Problem Does This Paper Address?
Wide Area Motion Imagery (WAMI) systems, often deployed on aerial platforms, capture vast amounts of data over large geographical areas. This data is crucial for applications like disaster relief, emergency response, and traffic management. However, the sheer volume of information, coupled with low frame rates and small target sizes in WAMI, presents a significant computational challenge. Traditional computer vision methods struggle with this scale and complexity, creating a bottleneck in extracting timely and meaningful insights. The paper addresses the need for a high-performance, real-time system capable of efficiently detecting and tracking multiple moving objects within these challenging WAMI datasets.
Methods
The core of the proposed solution is the GPU and Cloud Moving Target Tracking (GC-MTT) system. This architecture leverages a complementary approach:
- Cloud Infrastructure (Apache Hadoop): Used for distributed storage (HDFS) and parallel processing of large datasets. Hadoop's MapReduce framework distributes computational tasks across multiple nodes.
- Graphics Processing Units (GPUs): Employed for highly parallelized computations, significantly accelerating specific modules within the tracking pipeline. The Compute Unified Device Architecture (CUDA) is utilized for GPU programming.
- Hybrid Computation: The system strategically assigns tasks to either CPUs or GPUs based on their suitability. For instance, image registration, background/foreground generation, and data association are performed on GPUs for speed, while other tasks are managed by the Hadoop cluster.
- Block-Wise Image Registration: To handle the large image sizes and computational demands of registration, a fast, block-wise approach using CUDA is proposed. This involves SURF feature extraction, point matching, RANSAC for outlier removal, and transformation estimation to generate homography matrices.
- Front-End Web Server: A web-based interface allows users to interact with the system, configure parameters, and visualize results, providing a user-friendly way to access the processed data.
Datasets
The paper utilizes the following datasets for evaluation:
- Columbus Large Image Format (CLIF) dataset: Used for testing the speed-up performance of the GC-MTT system. The dataset consists of airborne images with a size of 2672 × 1200 pixels.
- AFRL CLIF dataset and WPAFB dataset: Mentioned and illustrated in Figure 1, these datasets are representative of WAMI imagery.
Key Insights and Ideas
- Synergistic Cloud-GPU Power: The paper demonstrates that combining Cloud and GPU computing offers a significant advantage over using either technology in isolation for WAMI processing. The distributed nature of the Cloud complements the parallel processing power of GPUs.
- Task-Specific Acceleration: By intelligently offloading computationally intensive tasks like image registration and foreground generation to GPUs, the system achieves substantial speedups.
- Real-Time Performance for Big Data: The GC-MTT architecture successfully transforms a computationally demanding problem (WAMI tracking) into a near real-time application, making the vast amounts of data generated by these systems more accessible for analysis.
- Extensibility: The proposed detection and tracking method is designed to be generalizable to other applications beyond WAMI, such as pedestrian tracking, group tracking, and Pattern of Life (PoL) analysis.
Key Results
- Significant Speedup: The GC-MTT system, particularly the Hadoop + GPUs configuration, achieved a speed-up of approximately 4 times compared to CPU-only processing. For instance, processing 56 frames took 90.94 seconds with Hadoop + GPUs, compared to 192 seconds with Hadoop + CPUs.
- Improved Frame Rates: The Hadoop + GPUs architecture achieved a frame rate of 0.61 frames per second, a notable improvement over Hadoop + CPUs (0.29 fps) and significantly better than traditional CPU or GPU-only approaches without Hadoop.
- Scalability with Nodes: Figures 13 and 14 show that increasing the number of computation nodes (from 1 to 4) in the Cloud architecture leads to improved average frame rates for both GPU and CPU-based processing.
- Reliable Detection: Figure 15 indicates that the detection performance of CPU and GPU is similar, with a trade-off between miss rate and false positive rate, suggesting a reliable detection mechanism.
Limitations and Future Work
- Limitations: The paper acknowledges that while the GC-MTT approach provides significant speedups, the overhead of Hadoop implementation might be more pronounced for smaller numbers of frames. The performance of CPU and GPU without Hadoop versions showed increased run times with larger frame counts, whereas Hadoop-based versions were more stable.
- Future Work: The authors suggest future research will focus on:
- Pattern of Life (PoL) trajectory analysis: Comparing PoL analysis with different cloud architectures for documenting and understanding target routines in urban areas.
- Anomaly Detection: Using the collected information to build models of normalcy and predict future actions, aiding in urban safety and intelligent highway systems.
- Further Optimization: Exploring ways to further reduce synchronization time and optimize resource allocation within the Cloud and GPU framework.
Summary updated