Real-Time Security Analytics For CCTV Cameras
1 year
vision artificial intelligence
media marketing
About Project
Security cameras are widely used in all types of organizations and for different purposes. Previously, humans had to monitor the recordings of those cameras to detect any unexpected event. This can’t continue in today’s world. At Odlica, we are revolutionizing video analytics by automating the monitoring process and providing a wide set of analytics. We built multiple machine learning models that can be separately integrated to surveillance systems. The built blocks include people counting, human path analysis, adherence to social distancing rules, panic detection, gender detection, face mask compliance, Received-attention assessment for a specific area, and multiple other models.
our role
artificial intelligence
UI/UX
Frontend Development
Backend Development
Client background
The system developed by Odlica is suitable for any company or organization that utilizes video surveillance cameras for public safety purposes. Companies providing video management systems can also utilize our ML models for state of the art analytics and configurability. Our AI and machine learning algorithms provide real-time analytics and insights, helping to improve safety and security in a variety of settings, including public spaces, companies, museums, and shopping centers.
Challenge
Developing accurate ML models for security cameras is challenging due to various factors such as manufacturer, resolution, positioning, and lighting conditions.
Security cameras streams are very different in the sense that they differ based on manufacturer, resolution, positioning, area covered, lighting conditions and many other aspects. Building accurate and robust ML models that perform with high standards given all mentioned factors is very callenging for many reasons.
- Providing real-time analytics without any significant latency
- The models should provide high quality results regardless of the environment, lighting conditions and with the lowest possible resolution
- The models output should be general enough to be easily integrated with different video management systems
- Ability to perform accurately in crowded environments where human detections might not be accurate.
Solution
We created a modular ML solution for camera analytics that uses a standardized input format, enabling quicker integration, individual model testing, and greater customization.
To build a comprehensive camera analytics ML solution, we decided to unify the input format for the ML models while building the solutions as building blocks that can be used separately or collectively. This allows for faster integration, ability to test the models in isolation, and better customization.
- A unified input&output scheme for straight forward integration with any platform.
- Making the ML models robust and more reliable by training for different cultures and environments
- Outputs are defined in a way that allow for easy visualization to enhance explainability
- ML models are independent of each other to allow for better "pay for what you use" model.
Outcome
Advanced Machine Learning Technology: Real-Time Analytics for Comprehensive Security Coverage
- 90-97% expected accuracy for ML models
- Integrated to +1K cameras so far
- Real-time analytics with minimal infrastructure costs
- Reducing dependence on human monitoring while providing higher accuracy