Project Overview
The Company:
Data analytics company serving government agencies
Industry:
Data Analytics
Services:
Video Annotation
Business Needs
The client needed to categorize and label vehicle movements (approach, turning, etc.) to feed machine learning solutions for government agencies. These solutions aimed to predict traffic issues, prevent accidents, improve road planning, and assist in civil engineering projects.
The client partnered with Hitech BPO to
- Label vehicle images based on predefined criteria to build training datasets to train machine learning models
- Evaluate video analytics for detecting queues, tracking stationary vehicles, and counting vehicles from live feeds.
The Challenges
- Recruiting annotation specialists proficient in standard automobile classification.
- Ensuring specialists had experience with complex computer vision models.
- Annotating images from videos with different lighting, weather, and traffic conditions required specialized skills.
- Organizing the workforce into shifts to handle large datasets efficiently.
- Choosing the right annotation method to accurately label data for machine recognition.
- Labeling images from unsteady or blurred traffic videos demanded highly skilled annotators.
- Managing the intricate process of annotating multi-frame video data.
- Counting pedestrians and bicycles from in-pavement loops not distinguished by direction (straight vs. turning right).
Solutions and Results
Want to annotate live traffic video stream?
Speak to us today- Identified, categorized, and labeled large volumes of vehicle and pedestrian images from traffic videos across major US and Canadian cities.
- Developed an efficient workflow for labeling after assessing pre-recorded and live traffic videos.
- Received video footage as pre-recorded videos and URLs to live streams.
- Uploaded pre-recorded videos to OneDrive, sorted by city.
- Accessed live videos via VPN to log into the city’s traffic camera network.
- Labeled and segmented data according to specific norms:
- Vehicles categorized by model, color, and direction.
- Objects classified into 14 categories including cars, SUVs, trucks, pedestrians, buses, vans, groups of people, bicycles, motorcycles, and traffic signals (green, yellow, red).
- Vehicles tagged and segmented by turning movement or direction of approach.
- Obstructed vehicles were not labeled.
- Ambiguous vehicles due to poor lighting or weather were re-validated by the client.
- Used line-based technique for accurate vehicle and object counts.
- Implemented status changes to track counts.
- Applied a red line to demarcate small, non-labeled vehicles.
- Senior auditors reviewed 10% of annotated images.
- Used data anomalies for training.
- Re-labeled any erroneous data.
- Organized labeled images on OneDrive by city.
- Generated reports detailing the number and types of annotated vehicles.
Business Impact
Millions of frames accurately labeled
Traffic dashboard for traffic management
Voluminous training data for ML Models
