Project Overview
An integrated data analytics company from San José’s, CA
Data Analytics
Business Needs
The client a data analytics company based in San José’s, CA providing solutions to government agencies in the energy, water, and communications arena was looking to categorize and label the huge number of vehicles based on movement like approach, turning movement, etc. The client needed to develop machine learning solutions to predict traffic-related issues like congestion, accident prevention, better road planning, lane movement, etc., and assist the Department of Civil and Environmental Engineering in developing proper road plans for smooth traffic movement.
Therefore, the client partnered with Hitech BPO to label vehicle images in pre-defined criteria that could be used to train the client’s machine learning models and evaluate if video analytics can detect queues, track stationary vehicles, and tabulate vehicle counts from live video feeds.
The Challenges
- Hiring annotation specialists who could understand the standard automobile classification.
- Specialists who had knowledge and experience in building complex computer vision models.
- Annotating images from videos under different lighting, weather and erratic traffic volumes required special skills.
- Dividing the workforce into shifts to train a massive amount of data.
- Selecting the right annotation method to accurately label the data for machines to recognize them visually, like a person.
- Skilled annotators with the ability to label images from unsteady or blurred traffic videos.
- Managing the complex process of annotating videos or multi-frame data.
- Counting pedestrians and bicycles from in-pavement loops not differentiated for directions of movement (going straight and turning right).
Solutions and Results
Want to annotate live traffic video stream?
Speak to us today- Huge volumes of vehicle and pedestrian images from live as well as historical traffic video feeds, from across major cities in the US and Canada were identified, categorized, and labeled.
- A workflow was planned after an initial assessment of vehicle and pedestrian images in pre-recorded traffic videos and live video streams, to expedite the labeling process.
- Input Data in the form of video footage were received as pre-recorded videos and URLs to live video streams.
- Pre-recorded videos were uploaded to OneDrive city wise.
- For live videos, human annotators securely accessed the VPN using pre-provided credentials to log into the City’s traffic camera network.
- Labeling and segmentation were done as per the following norms
- Vehicles were labeled by their category – model name, the color of the vehicle, and direction of the vehicle.
- Objects are classified into 14 categories – Car, SUV, small truck, medium truck, large truck, pedestrian, bus, van, group of people, bicycle, motorcycle, traffic signal-green, traffic signal yellow, and traffic signal-red.
- Vehicles were classified tagged and segmented by turning movement or by the direction of approach.
- Obstructed vehicles were not labeled.
- Any ambiguity on the vehicle due to poor light or weather conditions was re-validated by the client.
- Line-based technique was used to uniquely count vehicles and other objects.
- The state of the line changed from unoccupied to occupied and then back to unoccupied increasing the count of said line.
- A demarcation line (red line) was used to demarcate small vehicle that was not labeled.
- A team of senior auditors audited around 10% of the annotated images.
- Any anomalies or deviations in the data were used for training purposes.
- Any annotation found erroneous was taken up for re-labeling.
- City wise segregation helped in uploading the labeled images on OneDrive.
- Report listing number and types of vehicles annotated was generated for record purpose.
Business Impact
Millions of frames accurately labeled
Traffic dashboard for traffic management
Voluminous training data for ML Models
