Project Overview
Kephart needed a solution to automatically identify target individuals in raw video footage while minimizing manual intervention. Avahi developed a scalable, serverless system that ingests and preprocesses over 400 raw videos stored in AWS S3, dramatically reducing file sizes and processing times. The solution leverages AWS Rekognition for face and object detection, while AWS Elastic Transcoder, Lambda, and DynamoDB handle video optimization and metadata extraction. As a result, the system delivers rapid, cost-effective searches and improved operational responsiveness.
About the Customer
Kephart is a forward-thinking leader in the video analytics and security space. Focused on enhancing surveillance and content analysis, the company relies on innovative technologies to deliver real-time insights that drive safety and operational efficiency.
The Problem
Kephart faced the challenge of processing large volumes of raw video footage to identify a specific target individual—a task that was both time-consuming and error-prone when done manually. The need to extract only the relevant segments containing human faces and associated metadata, such as timestamps and confidence scores, was critical. Without automation, manual processing would have led to delays, increased costs, and compromised security outcomes.
Why AWS
AWS was chosen for its robust suite of AI and machine learning services that perfectly matched the project’s requirements. AWS Rekognition provided accurate face and object detection out of the box, while AWS S3, Lambda, and DynamoDB ensured a scalable, secure, and cost-effective solution. The cohesive AWS ecosystem allowed for seamless integration and rapid deployment, meeting the high-volume,performance-driven needs of the project.
Why Kephart Chose Avahi
Kephart engaged Avahi because of our extensive expertise in designing and deploying AWS-based AI solutions. Our proven track record in building scalable, efficient cloud architectures gave the customer confidence in our ability to tackle complex video processing challenges. Avahi’s innovative approach and deep understanding of AWS services ensured that the solution would not only meet but exceed performance and operational efficiency goals.
Solution
Avahi architected a comprehensive system that begins by ingesting over 400 raw videos stored in AWS S3. The initial phase utilizes AWS Elastic Transcoder to preprocess the videos, significantly reducing file sizes— from a 513MB video down to 4.3MB—and slashing processing times from 5 minutes to 13 seconds.
AWS Rekognition is then employed to detect faces and objects within the processed videos, extracting key metadata such as timestamps, landmarks, and confidence scores. This metadata is stored in AWS DynamoDB, enabling rapid and efficient querying.
A serverless API, powered by AWS Lambda and API Gateway, was built to allow users to upload a target image and initiate a search across the video metadata. Additionally, AWS Glue automates further data processing to ensure comprehensive results when a target image is not initially found.
Together, these integrated AWS services form a streamlined pipeline that optimizes video storage, accelerates data processing, and delivers rapid identification of target individuals.
Key Deliverables
- Preprocessed video files optimized using AWS Elastic Transcoder
- Automated video processing pipeline integrating AWS Rekognition for face & object detection
- Extracted metadata stored in AWS DynamoDB
- Serverless API for target image upload and search via AWS Lambda and API Gateway
- Extended data processing using AWS Glue
Project Impact
The solution dramatically improved processing efficiency and search accuracy, eliminating labor-intensive manual tasks. By reducing video file sizes from 513MB to 4.3MB and processing times from 5 minutes to 13 seconds, the system lowered storage costs and accelerated the identification process, enhancing overall security responsiveness.
Metrics
- Video size reduced from 513 MB to 4.3 MB
- Processing time decreased from 5 minutes to 13 seconds
