AI-Based Parking Surveillance
Client’s company provides innovative parking technology solutions designed to enhance safety and efficiency in parking management so the client came to us to build an AI-based solution that monitors and enforces parking rules across the US. Helps to reduce costs, wrong fines, manual human labor and increased processing speed.
Project Challenges
Client’s company provides innovative parking technology solutions designed to enhance safety and efficiency in parking management so the client came to us to build an AI-based solution that monitors and enforces parking rules across the US. Helps to reduce costs, wrong fines, manual human labor and increased processing speed.
Different conditions for obtaining images of cars in a parking lot. Images of cars for comparison can be obtained at different times of day and in different weather and illumination conditions. Distinguishing cars can be complicated by the presence of rain, shadows, illumination from car headlights, etc.
Insufficient amount of data for training neural networks. Some cases that need to be detected are relatively rare, which makes it difficult to obtain enough data for training. This requires the generation and use of synthetic data that will reproduce the required situations as realistically as possible.
High requirements for accuracy. In order for the developed system to be truly effective, high requirements are imposed on the accuracy of detection of the required objects and situations in the parking lot. No less important is the ability to identify cases in which there is insufficient confidence in the solutions offered by the AI and manual validation by an expert is required.
Solution
We managed to create universal solution for different facilities that caters the decision process whether to approve or reject the violation according to the established rules. Our system validates violation using a series of images and some metadata. Our system uses AI to:
Detect the presence of a parked car in a parking space
Compare cars on the entrance and violation images, as well as license plates, if they are readable
Detect car movement on a series of entrance images
Detect license plates which are unreadable due to occlusions, image quality, cropped license plates, etc.
Detect defective images that are not suitable for obtaining any information: with glitches, artifacts, heavily backlit or dark.
-
Detect the presence of a parked car in a parking space
-
Compare cars on the entrance and violation images, as well as license plates, if they are readable
-
Detect car movement on a series of entrance images
-
Detect license plates which are unreadable due to occlusions, image quality, cropped license plates, etc.
-
Detect defective images that are not suitable for obtaining any information: with glitches, artifacts, heavily backlit or dark.
Tech Stack
Python
TensorFlow
OpenCV
aiohttp
RabbitMQ
Postgres
Keras
XGBoost
AWS
CloudFormation
Results
This allowed the client to implement it at their sites and automate the manual work of their staff, which in turn reduced business costs. Also the result of our work was a library for automatic additional training of models (data load, data generation, training and evaluation processes, data and prediction visualization, annotation process description, datasets signature description) to facilitate further development of the project and allows to improve the product yourself, without the intervention of developers.
As a result we successfully created a solution that showed 95% accuracy in determining the outcome.
