Contact

Intelligent vehicle counter

Intelligent vehicle counter

About the client

Our client is a private company that deals with counting vehicles on the road.

Project info

Challenge

 

The Client is engaged in providing vehicle traffic counting for government services based on recorded video manual analysis. They considered applying an automation tool for vehicle detection and counting that would recognize a vehicle type according to certain categories among the general road traffic. This tool would optimize human efforts, eliminate errors, and speed up the workflow and consequently the results.

 

Solution

The Avenga team developed a semi-automated system for road traffic load estimations. The application is designed to expose a web-based user interface that allows traffic video processing, reporting, and validation. Our data science team was responsible for developing the machine learning model based on computer vision technology. The solution consisted of detecting and classifying a vehicle on a video stream on multiple categories, tracking the vehicle through each frame and eventually counting it using a deep learning approach.

 

During the first phase of the project, the pre-trained YOLO architecture was leveraged on the COCO dataset which was able to detect a limited number of required vehicle categories. Later development phases of the project intended to retrain the customized YOLO neural network in order to correctly classify and count all the required types of vehicles, which was impossible in the beginning due to the absence of labeled data. We selected an approach that immediately allowed us to process video streams and produce a labeled dataset for future needs simultaneously. That solution consisted of a smart feature with the ability to produce a customized labeled traffic images dataset which should enable the training of the neural network to detect vehicles on predefined categories from scratch, in later development phases.

 

After detecting the bounding box of the vehicle using YOLO, we utilized a vehicle tracking system based on the SiamMask approach. The results of the tracker were additionally analyzed in order to specify the direction of movement (left and right) and to enhance the tracker algorithm.

 

The Avenga team also developed a cutting-edge automated algorithm for DROI and counting line detection. This allowed for fully automated video processing without manual positioning of counting lines for each video recording, improved the model performance and lowered costs.

 

Besides developing an algorithm for counting vehicles, we built the solution for video quality checking and sorting videos to ‘manual’ or ‘automated’ buckets leveraging Azure Functions service.

Results

With the use of innovative tech, we contributed to the “smarter” use of an integral part of modern cities – the transport networks and traffic flow measurement and maximized the value of visual data by implementing cutting edge algorithms and models for image and video data analysis and manipulation.

The client benefited from:

 

  • Implemented scalable semi-automated vehicle categorization and counting involving computer vision techniques to make the process quicker, more accurate, and less resource-consuming.
  • Customized algorithms for predefined vehicle categories and edge cases
  • Enabled AI-assisted human workflows for better efficiency and scalability

Technology used

  • Python

  • OpenCV

  • Azure

  • YOLO

  • SiamMask

Challenge us
We’d like to hear from you. Please use the contact form below and we’ll get back to you shortly.