How Softacom Cut CCTV False Alarms by 99.9% with AI

false alarm per camera per week – down from 1 every 10 minutes

12+

neural models were used for experiments

32K

USD saved monthly

working hours saved each week

Description

Our customer
Softacom was contacted by a security systems provider from Texas, USA. The company specializes in the development, consulting, and installation of fixed video surveillance systems (CCTV) for buildings and facilities that host public events such as concerts, football matches, and more.

The company offers their own video surveillance platform, which unifies cameras from multiple manufacturers into a single integrated system. Among the video cameras used are such as Hikvision, Axis, Avigilon, Panasonic, and others. Many of their clients, including government institutions, required AI solutions that could be deployed entirely on-premise without relying on third-party cloud services or subscription-based models.

Although almost every manufacturer mentioned above offers its own turnkey video surveillance management platform, our client has been successfully implementing their own solution for over 30 years. It meets the needs of their customers by providing classic video surveillance functions. These functions include detection, recording, and the organization of control and monitoring centers, as well as deep integration with other security subsystems, such as access control.

The Customer’s Request

The company wanted a solution that would increase the accuracy of detecting people and vehicles and reduce the number of false alarms by 50% when such objects are detected outside designated attendee areas. For example, detecting people at night in service zones, loading and unloading areas, VIP boxes, and similar locations.

The security teams of organizations – our client’s customers – regularly complained about false alarms and detections, which reduced the overall level of site security. Due to the high number of false triggers, the system was no longer taken seriously, it overwhelmed the staff, and as a result, the required functionality simply didn’t work.

There were situations, especially during poor weather conditions, when security personnel received up to five false alarms per minute. These led them to lower the system’s sensitivity, which in turn led to a lack of trust in the entire solution.

An important requirement was to avoid using third-party services to eliminate subscriptions and monthly costs. Some of our client’s customers include government organizations with fixed budgets. Based on the results of the tender process, they are ready to buy equipment outright but are not prepared to pay monthly fees to companies like OpenAI, Google, or Microsoft for the use of their AI solutions.

Our task was to develop a Proof of Concept (PoC) solution, test it at a real facility, and – if successful – integrate it into the existing system as a separate add-on with a pricing model.

Struggling with false alarms or AI model integration? Let’s talk about computer vision that really works.

Learn more

Challenges

Reducing false alarms in high-traffic surveillance zones, such as stadiums or arenas. A high number of false alarms led to security teams losing trust in the system.
Avoid using third-party cloud subscriptions due to client preferences.
The existing system lacked sufficient computational resources for deep learning-based detection.
Enhancing the client’s existing Delphi-based desktop software to work with YOLO models.
Balancing detection speed and accuracy while maintaining low hardware overhead.

Solutions

Analyzing the System and Developing a Modernization Approach

After analyzing the existing system, we created a roadmap for implementing the future solution. The current system receives detection event decisions from the SDK of Hikvision systems. We assumed that these systems were based on mathematical algorithms, possibly using libraries like OpenCV. This approach can be considered a classic one, and it has its pros and cons.

We decided to develop a system that processes the video stream from cameras using GStreamer, extracts a sequence of static images, and analyzes those images using a neural network. This analysis includes both pre- and post-processing using custom algorithms developed by Softacom AI Lab (we will describe them in more detail later).

Model Selection and Performance Requirements

The first step on our side was to select an optimal neural network model that would already be pre-trained to detect people and vehicles, as the client did not want to invest in training neural networks themselves. We already had experience working with both commercially licensed models and open-source ones.

Together with the client, we defined the system requirements – image resolution for analysis, processing speed (1 frame per 100-500 milliseconds), and the maximum number of video cameras supported. We also requested sample recordings from the client’s video cameras, which we would use later to check the quality of the new solution. These recordings included metadata that allowed us to track the number and timing of both false and accurate detections.

Challenge: Overcoming Resource Constraints at Scale

The challenge was figuring out how, on the client’s side and in real-world environments, we could process images from 100 video cameras at a rate of 1 frame every 100 or even 500 milliseconds, especially considering that the existing system did not have the required computational resources.

Based on future installations, the choice of development technology and runtime environment also became an important consideration. We evaluated native programming languages such as C++, .NET, Delphi, and Python.

We conducted experiments with 12+ models (including some sourced from Hugging Face). In parallel, we tested a multimodal image verification approach using Softacom AI Lab.

Multimodal processing means that the image is not analyzed by a single model. Instead, it goes through sequential processing by several models, each responsible for a specific detection task. The final result is then formed either by an additional model or through a programmed algorithm.

Based on the results, we chose YOLO models, specifically YOLOv11. After that, the client directly contacted the model’s rights holder to discuss licensing for commercial use.

Custom Runtime Integration and Open-Source Tools

As one of the possible execution environments for the solution, we developed an architecture based on RACK, a specialized computer designed to run DeepML models. Our client offers this RACK as part of their solution for high-quality event detection for video surveillance systems.

Also, to enable integration with client’s existing desktop software used by operators for event detection, we made enhancements to their application written in Delphi. Since we could not find a ready-made solution that would allow working with YOLO models directly from Object Pascal code, we developed our own solution.

We developed a custom wrapper class for working with YOLO models from Object Pascal code via Python libraries. Our wrapper class is available as open-source on GitHub: https://github.com/SoftacomCompany/Delphi-YOLO-ONNX-RuntimeWrapper.

Challenge: Reducing False Alarms During Poor Weather Conditions

Reducing false alarms during poor weather conditions was challenging. Rain, fog, or snow caused misdetections because of moving debris and shadows.

To adapt the system for bad weather, we used OpenCV filters, Gaussian Mixture Models, and CLAHE. The system learnt to see through weather noise. Synthetic weather data retrain YOLOv11 to handle these conditions better, while a secondary model handles tricky cases, raising alarms only when both models agree. So, detections must persist across multiple frames to be valid, and sensitivity adjusts based on time and weather data, all while optimized hardware ensures real-time processing across 100 cameras.

This approach allowed us to reduce false alarms during poor weather conditions from 5 per minute to 1-2 per minute, which accounted for up to 80%.

AI shouldn’t just be powerful – it should be practical. See how we delivered both.

Talk to Our Experts

Outcome

Thanks to our expertise in Computer Vision and Machine Learning, we designed and integrated a custom module for detecting people and vehicles using DeepML. This module was successfully deployed across more than 100 surveillance cameras, significantly improving detection quality by 99,9% and meeting their end clients’ strict data privacy (raw video footage, detection alert metadata and logs, AI model inputs and outputs, annotated images, and more) and budgetary requirements.

Detection accuracy significantly improved: the average number of false alarms dropped from 1 per camera every 10 minutes to just 1 per 1 camera per week, which is 99,9%. The Softacom team managed to reduce the number of false alarms by 99.9%, exceeding the requested 50% reduction. During poor weather conditions, false alarms also dropped by 80%, from 5 per minute to 1-2 per minute.
The customer support team stopped spending up to 40 working hours per week handling complaints and tickets related to system performance, saving over $32,000 per month.
The number of tickets caused by false alarms dropped by 92%. The overall feedback from the on-site security personnel was positive. Employees regained trust in the system.
Our client now has a more reliable system.

Save 40 hours/week and $30K/month – like our client did.

Save