31
Jul
Real-time surveillance and advanced tracking capabilities are more crucial than ever for national security. A leading public defense company leveraged MLtwist’s technology to enhance its ability to accurately track people and vehicles in high-frame-rate, low quality image videos, detecting, tracking, and annotating human and vehicle movements across tens of thousands of frames per video. This required processing videos at 29 frames per second and precise bounding boxes on every single frame, driven by AI that performed well enough to accurately interpret low-quality images The challenge was immense and MLtwist’s cutting-edge AI-driven data labeling solutions provided the accuracy and efficiency needed to meet it head-on.
The task for this company: processing low quality videos at 29 frames per second while applying precise bounding boxes on every single frame. Traditional methods for transforming raw video data proved inefficient, often failing to maintain consistency across frames or struggling with edge cases such as occlusions, motion blur, edge of frame, low resolution, and varying lighting conditions.
While AI performs well on high-quality video, AI applied to low-quality video may look right on the surface and end up being wrong on the frame by frame review. Providing a technical solution that was high quality and scalable was a challenge, but MLtwist’s cutting-edge AI-driven data labeling solutions provided the accuracy and speed needed to meet it head-on.
Automated Pre-Labeling:
MLtwist deployed a recently released AI model to pre-label each frame, identifying people and vehicles with bounding boxes while tracking their movements across sequences. The quality of the pre-labeling was not high enough to gain efficiency and in some cases it actually proved counter-productive, increasing the time per task for the human in the loop with no gain in quality output. So MLtwist evolved the AI pre-labeling for this particular use case, thanks to a flexible technology that adapts its flow for each data need.
Data Cleaning & Transformation:
The pre-labeled data was refined using proprietary algorithms to remove noise, correct misalignments, and ensure the highest level of consistency across frames.
High-Accuracy Bounding Boxes:
MLtwist brought in a labeling partner to ensure that bounding boxes were not only tightly fitted but also dynamically adjusted frame by frame to maintain precision, even in complex scenarios like overlapping objects, fast movements, and partial obstructions.
Quality Assurance (QA):
A rigorous QA process was implemented, with both automated validation checks and human reviewers ensuring that annotation accuracy exceeded industry standards. These validation checks were reports that examined the raw results and reported on anomalies, which often were the result of a mis-click or applying the wrong label.
Formatted Output for Seamless Integration:
The processed and labeled video data was packaged in a structured JSON COCO format with additional metadata, dubbed COCO+, allowing the defense company to seamlessly integrate it into their analytical and surveillance systems.
By leveraging MLtwist’s capabilities, the public defense company significantly improved the efficiency and accuracy of their tracking operations. Key benefits included:
MLtwist’s AI-powered data processing is changing how defense organizations analyze video for public defense applications, delivering precise, high-quality tracking at scale. By delivering high-precision bounding boxes with unparalleled accuracy and automating the cleaning, transformation, labeling, and QA of high-frame-rate footage, MLtwist is helping defense organizations stay ahead in the rapidly evolving landscape of security and surveillance with scalable, high-precision solutions.