Task
The client approached us with an interesting challenge: to annotate surveillance footage from a factory entrance to enable automatic employee identification and ID matching with the access control system. The video contained footage from three different angles: two cameras inside the entrance area and one monitoring the exit.
This project presented several unique challenges that required creative problem-solving and workflow adjustments:
- The initial data volume significantly slowed down processing.
- The neural network’s pre-annotation had inaccuracies, requiring manual refinement.
- Introducing a frame-cutting stage extended the timeline but substantially improved the final quality.
Solution
-
- 01
-
Preprocessing the Video:
- Filtering Footage: Up to 80% of frames contained irrelevant data, so we implemented a video-cutting stage.
- Optimizing Data Volume: After processing, the total number of frames was drastically reduced.
-
- 02
-
Data Annotation:
- Neural Network Pre-Annotation: Automatic detection of people in the footage.
- Manual Refinement: Removal of false detections and precise object boundary adjustments.
-
- 03
-
Automated ID Matching:
- Developed a script to automatically match employee IDs with the annotated footage.
-
- 04
-
Validation & Quality Control:
- Verified pre-annotation accuracy and corrected errors.
- Final validation with object boundary refinements to enhance precision.
Results
Optimized workflow reduced the frame count from 50-60 thousand to 8 thousand per dataset.
Neural network pre-annotation accelerated the annotation process.
Improved annotation quality through thorough filtering and precise object boundary adjustments.
