Contact Us
CLIENT SUCCESS STORY

Enhancing Drone AI Performance for a Surveillance Company through Video Annotation Services

The Client

A Leading Security Firm in the US

This company specializes in providing drone-based aerial surveillance and security solutions tailored to the needs of businesses in agriculture, construction, and real estate. Leveraging high-resolution imaging and real-time data collection, they enable businesses to gain actionable insights, enhancing decision-making and operational efficiency in their respective industries.

Project Requirements

Image Labeling for Drone Footage

The client has developed a proprietary AI system to detect and track drones or other objects of interest captured via drone footage in various challenging environments. They were focused on optimizing this AI, specifically its ability to distinguish between drones in flight as well as identify and track drone movements precisely (at varying altitudes, in diverse lighting scenarios, and across all stages of flight).

For this purpose, they needed accurate video annotation services. Our team was provided drone footage (captured using both standard and infrared cameras). The footage included recordings taken in various conditions—daylight, nighttime, low visibility, and during high-speed drone movements.

Project Challanges

Complexities of Infrared Footage, Visibility, and Erratic Drone Movements

The footage provided to our video labeling team was 55 hours long (approximately 100,000 frames). However, annotating this footage involved certain challenges:

  • Varying Thermal Signatures- The footage included data from infrared cameras, with fluctuating thermal signatures that made it difficult to consistently identify drones and distinguish them from other objects, especially in areas with mixed heat sources.
  • Poor Visibility In Low-Light Conditions- A significant portion of the videos was captured at night or in low-light environments, making object detection in drone footage more challenging.
  • Unpredictable Drone Movements- The drones being monitored exhibited rapid and unpredictable movements, often at high speeds, which made it tough to capture their precise flight paths without missing any key frames.
  • Long-Range Shots- Some recordings included drones captured from significant distances, leading to reduced clarity and detail. This required our team to carefully adjust annotations to account for varying object sizes and ensure accurate identification.

Bounding Box Video Annotation with Multi-Level Validation

A dedicated team of twenty data annotation experts was assigned to this project. As per client instructions, we worked on CVAT- a web-based, open-source image and video annotation tool originally developed by Intel.

Here’s the human-in-the-loop video annotation approach we adopted for this project:

  • Leveraging Automated Annotation where Feasible: To accelerate the annotation process while maintaining accuracy, we implemented semi-automated bounding box suggestions for straightforward frames, with manual verification and corrections where needed.
  • Enhancing Frame Opacity for Infrared Footage: To tackle fluctuating thermal signatures, our team adjusted frame opacity, allowing for better visibility of drones against mixed heat sources. This enabled more consistent identification and differentiation of drones from other objects.
  • Compensating for Poor Visibility: For footage captured at night or in low-light scenarios, we applied contrast adjustments and brightness enhancement techniques within CVAT. This made it easier for annotators to accurately identify drones and ensure precise labeling.
  • Manual Adjustments for Unpredictable Drone Movements: Given the rapid and erratic movements of drones, we employed manual frame-by-frame adjustments to accurately capture their flight paths. Annotators used real-time frame scrubbing and playback to ensure bounding boxes tracked drones seamlessly, even during high-speed movements.
  • Optimizing Annotations for Long-Range Shots: To address reduced clarity in long-range footage, we used dynamic zooming within CVAT, enabling annotators to focus on distant drones without losing detail. This helped maintain accurate bounding box placements, even when drones appeared as small objects.
  • Multi-Level Annotation Validation: Every annotated segment underwent multiple rounds of review, including peer validation and supervisor oversight. This ensured high accuracy and consistency across all 100,000 frames, minimizing potential errors.
  • Iterative Feedback Cycle with the Client: We shared batches of annotated data with the client for review. This allowed for adjustments based on their evolving needs and ensured that the final dataset perfectly aligned with their AI model training objectives.
  • Custom Annotation Templates: For repetitive drone detection scenarios, we created annotation templates to standardize the process. This improved efficiency while ensuring consistency across similar footage segments.

Project Outcome

project outcome
Enhanced the client’s drone surveillance system with an impressive 30% higher accuracy.
project outcome
Accelerated AI response time with 20% surge in overall operational performance.
project outcome
Better tracking accuracy and expanded AI detection capabilities with near-zero false positives.