Overview
Through collaboration, innovation, and a data-driven approach, Sinenza empowered our
client to transform their traffic analysis processes, yielding substantial operational and
cost efficiencies.
The Customer
Our client, a leading provider of traffic, transport, and mobility data solutions,
specializes in supporting strategic planning, traffic impact assessments, property
management, and retail evaluations. Their focus extends to optimizing existing transport
management systems to enhance overall efficiency.
The Challenge
Our engagement with the client centered around addressing a critical issue - accurate
traffic counting. The primary objective was to identify and tally passing vehicles on
non-real-time street camera surveillance videos, categorizing them by type (car, truck, bus,
etc.). Subsequently, the project expanded to include license plate recognition.
Initial scenario
Prior to our intervention, the client relied on a manual approach to vehicle counting. They
had developed a rudimentary software tool that utilized video input and a game controller
for efficient counting and categorization of vehicles. Although this tool expedited the
counting process, human involvement remained necessary. Despite efforts to streamline the
process and outsource counting tasks to low-cost workers, such as those in India, the
profitability of this approach was limited. Demand grew as new customers were acquired, but
challenges persisted.
Decision to innovate
Recognizing the potential for automation, the client explored the development of advanced
counting software. Cloud computing costs were projected to be significantly lower than human
labor expenses. Furthermore, automation could deliver unparalleled accuracy, overcoming the
shortcomings of manual counting. The pivotal concern centered on the investment required for
development and the anticipated returns over time.
Why Sinenza?
Our team's expertise in image recognition and neural network development positioned us as
the ideal partner for this project. We provided transparent insights into the functionality
of AI algorithms, detailing client involvement such as data collection and tool training. We
proposed a comprehensive workflow encompassing vehicle identification, tracking, and
fine-tuning of output. Our well-defined accuracy metrics offered confidence and enabled easy
comparison with human performance. We presented an adaptable, fully automated processing
architecture on the Azure cloud platform.
Our Solution
Our initial focus was extensive research to identify cutting-edge solutions for object
detection and tracking. After customizing these solutions to our needs, we embarked on
training data generation. We developed an intuitive data labeler tool and collaborated with
the client to outsource labeling to their manual counting workforce. While the training
dataset's quantity was satisfactory, the quality suffered due to the workforce's prior focus
on simple counting rather than precise labeling.
Synthetic Data Solution
Recognizing the limitations, we adopted a synthetic data approach. Leveraging a manipulable
computer game, we devised a processing flow to extract training data from within the render
pipeline. This approach yielded a substantial volume of high-quality data, vastly surpassing
our performance goals.
Refined User Interface
Simultaneously, we developed a user-friendly interface empowering the client to fine-tune
the processing flow for specific scenarios, such as night or inclement weather conditions.
Cloud Deployment
and Expansion
Each stage of our pipeline found its place in the cloud, enabling an end-to-end solution to
automatically process camera surveillance videos. Scalability and cost-effectiveness were
key priorities. In a subsequent phase, we extended the solution to include license plate
recognition under optimal visibility conditions and high-resolution videos. Our previously
established data-gathering mechanism proved instrumental in meeting these new requirements.
Results
With our solution in place, the client achieved successful processing for 70% of their
surveillance videos without requiring manual intervention. An additional 85% of videos
necessitated only minimal refinement. This translated to a significant 70% reduction in
recurring processing costs, achieved by minimizing reliance on manual labor. Remarkably, the
development costs were fully recouped within a year of implementation.