Distributed query processing for mobile surveillance

Stewart Greenhill, Svetha VenkateshACM Multimedia 2007 · DOI

TL;DR

This paper introduces HORUS, a novel architecture for distributed query processing in mobile surveillance systems. It tackles the challenge of retrieving and analyzing vast amounts of semi-permanent, multi-modal data from networks with limited and unreliable connectivity, enabling adaptive quality-of-service for surveillance applications.

What Problem Does This Paper Address?

Traditional surveillance systems rely on static camera networks, which offer limited spatial coverage. The advent of mobile surveillance, using cameras mounted on vehicles like buses, significantly expands coverage but introduces substantial challenges. These include:

  • Data Volume and Distribution: Mobile platforms generate massive amounts of data (video, GPS) that are distributed across numerous mobile sensors and static infrastructure.
  • Limited and Unreliable Connectivity: Communication links between mobile sensors and data centers are often intermittent, low-bandwidth, and scheduled (e.g., only when a bus returns to its depot). This makes continuous data retrieval impossible.
  • Semi-Permanent Data: Data on mobile devices is often stored temporarily and needs to be re-used, meaning it's only available for a limited time before being overwritten.
  • Complex Querying Needs: Surveillance queries go beyond simple database lookups. They require spatio-temporal constraints, the ability to specify varying levels of detail (sampling rates), and the integration of multi-modal data (video, GPS).

Existing distributed query processing systems often assume stable, high-bandwidth connections and don't adequately address these specific constraints of mobile surveillance.

Methods

The HORUS system is built around a three-tier architecture:

  1. Application Server (A-Node): The central point for clients, holding aggregated data and managing high-level queries.
  2. Depots (D-Nodes): Intermediate storage points, typically at bus depots, that collect data from mobile sensors.
  3. Sensors (S-Nodes): The mobile cameras and GPS units on vehicles.

Key methodological contributions include:

  • Adaptive Query Processing: The system dynamically adjusts query execution and data retrieval based on available network bandwidth and connectivity. It prioritizes high-value queries (e.g., specific observer viewpoints) over lower-priority background coverage data when bandwidth is scarce.
  • Novel Query Operators: HORUS introduces "coverage operators" for low-detail, broad spatial area queries and generalizes "observation operators" for high-detail queries over smaller areas.
  • Distributed Storage and Retrieval: A distributed storage system is employed, with data progressively moved from mobile sensors to depots and then to the application server. Retrieval is "on-demand," meaning data is fetched only when needed for a query.
  • Query Resolution and Materialization:
    • Query Resolution: This process maps spatial queries and trajectory data to specific video segments. It involves functions like resolveQuery (for visualization) and resolveCamera (for data gathering).
  • Materialization: The system determines which data needs to be retrieved and stored, prioritizing based on query type and available resources.

Datasets

The paper uses several datasets for evaluation:

  • CATLIVE: Live output from approximately 20 buses in Perth's Central Area Transit (CAT) service. Data was collected online with about a one-day latency.
  • CATVID: A 1.2TB dataset of video and trajectory data from the same 20 buses over approximately 4 days, collected directly from bus recorders.
  • CATGPS: Historical GPS data for these buses, spanning up to 9 months. This dataset was used to simulate network conditions and test aspects not fully achievable with the live CATLIVE data due to infrastructure limitations.

Key Insights and Ideas

  • Prioritization is Key: In bandwidth-constrained environments, a strict prioritization scheme (e.g., observer queries over coverage queries) is essential for delivering meaningful results. The system effectively sheds lower-priority data to preserve higher-priority information.
  • Adaptive Quality of Service: The system's ability to adapt its data retrieval and processing based on real-time network conditions is crucial for its robustness. This allows it to provide the "best possible level of detail" given the constraints.
  • Bridging Static and Mobile Data: The architecture effectively integrates data from both static (depot) and mobile (vehicle) platforms, creating a unified surveillance view.
  • The Value of Coverage Data: While observer queries are high-value, coverage data provides essential background context for browsing and understanding the broader scene, even if at a lower resolution.

Key Results

  • The system successfully demonstrated its ability to resolve individual queries and cameras by merging contributions from multiple mobile sensors.
  • Simulations using the CATGPS dataset showed that even with limited bandwidth (e.g., 250 Kbit/s up-links), the system could prioritize and deliver high-priority observation data (retaining 97% of L1 data) while adapting by dropping lower-priority coverage data (L3 data reduced to 6%).
  • The system effectively manages data retrieval from sensors before their storage is reclaimed, a critical aspect for mobile platforms.

Limitations and Future Work

  • Proprietary Depot Infrastructure: The authors noted limitations in the existing proprietary depot infrastructure (e.g., inefficient scheduling, fixed download time slots, inability to select specific channels or resample video). They expect these to be addressed by vendors in the future, enabling more comprehensive testing.
  • Response Time for Interactive Queries: The response time for interactive queries is dependent on data location and query complexity, which is difficult to characterize without specific scenarios. This is identified as a subject for future work.
  • Full-Scale Testing: While the system was demonstrated on a real bus network, full-scale testing with a larger number of vehicles and more diverse network conditions is a natural next step.

Summary updated

BibTeX
@inproceedings{greenhill2007distributed,
  author = {Stewart Greenhill and Svetha Venkatesh},
  title = {{Distributed query processing for mobile surveillance}},
  booktitle = {ACM Multimedia},
  year = {2007},
  doi = {10.1145/1291233.1291331},
}

Loading related content...

Figure