How deep learning techologies are making cities safer


Video systems today have become the foundation of city-wide security, but these systems already produce more video than cities can use. A constant problem is the search for personnel to monitor images from video cameras and search archives. And increasing number of cameras will only make this problem worse.


Artificial Intelligence (AI) can help solve this problem by simplifying the search and analysis of video streams and reducing the need for human resources. But conventional AI is difficult to deploy, often costly and time consuming.



“It can take four to six months for the development team to customize the AI algorithm for new goals or locations,” said Sean Lin, product manager for GeoVision Inc. “And the results can be disappointing, with too many false alarms and other errors,” He emphasized that “what cities need is an easier way for operators to pinpoint what they are looking for in important video streams, instead of searching for needles in haystack. "


The advent of solutions based on deep learning technology significantly improves computer vision systems and facilitates video analytics. Such systems are more powerful, easier to deploy and available today.


Thanks to deep learning, different models can be trained in accordance with the conditions of the environment in which the cameras are installed. Algorithms adapt to each situation without the need for further changes.


Huge amounts of video in this case are an advantage, not a burden. Deep learning systems can absorb data to adapt to new conditions and requirements.


Deep learning changes the rules of the game

Thanks to deep learning, computer vision technologies such as face recognition or motion detection are becoming more advanced, changing video surveillance and other security areas.



In a controlled environment, traditional algorithms are doing ok, but they are usually created for specific conditions of use. Finding an object or person crossing a given virtual line is basically a simple yes-no algorithm. Problems arise when similar algorithms are used in more complex circumstances.


Lin gives simple examples: “When you take the traditional algorithm and use it for different video surveillance locations — some cameras can be placed in the park, others on the street — the corresponding locations are displayed differently on the video streams. Traditional algorithms will not be able to cope with such differences. ”


“On a busy street, a large number of false alarms of a motion detection system or an intruder alarm due to the constant movement of people are possible. In this case, traditional algorithms show the limit of their capabilities, ”he said.


Another common scenario is face recognition when the police identified the wanted person. “Thanks to deep training, we can add this person’s face to the database with just one image or video from the database. After that, our software will be able to automatically analyze all the video surveillance records for a month, two months and find this person, ”Lin said.


He also predicts that in the future it will be possible to use even a sketch instead of a photograph. Although the recognition accuracy will decrease, this option would be impossible using traditional algorithms.


This is where the GeoVision Smart Video Management Solution (GV-VMS) comes in, which improves the AI model and allows for more complex and intensive analysis. GeoVision deep learning algorithms can be tailored to a wide range of conditions, including:

  • Counting people or objects moving in two directions
  • Detection and recognition of faces for various purposes
  • Masking faces to ensure privacy in the video
  • Anti-fog video in poor visibility for clarity
  • Stitch video from multiple cameras into one panoramic view
  • Stabilize video in a vibrating environment
  • Counting people in a crowd with attendance restrictions
  • Eliminate distortion caused by wide-angle lenses
  • Smart event search for motion detection in the surveillance area

Complex solution


The unique GeoVision deep learning functionality is based on an integrated system consisting of cameras, recording servers and a video control center. It connects with GeoVision IP-cameras and third-party IP cameras using standard protocols, as shown in Figure 1. This is possible thanks to Intel® processors that increase video processing efficiency and deep learning capabilities.


Figure 1: Smart Video Management System (VMS) GeoVision


Based on Intel® x86 Architecture, Server GV-VMS takes full advantage of the Intel® Core ™ processor. Using Intel® OpenVINO ™ Toolkit enhances the performance of deep learning video analytics by eight to ten times. This provides more capacity for simultaneous video processing without any additional requirements.


Geovision cameras have advanced deep learning capabilities. Cameras can send notifications when they detect, instead of sending all the videos to the central station for analysis, which reduces the delay before taking action.


Most cities also have outdated video systems with cameras, gateways, and software. GeoVision Application Programming Interfaces (APIs) and Software Development Kit (SDK) provide connectivity to old hardware and software. GeoVision Control Center provides unified cloud-based management software that integrates all IP cameras into a common security and management system


For example, the Vatican has been using video surveillance systems for decades. Over the years, this has led to the accumulation of different cameras, gateways and software tools from different suppliers. Working with GeoVision, the Vatican strategically integrated old cameras and software into a centralized surveillance solution. GeoVision solution allows you to create a unified system for monitoring video streams at 140 sites throughout Rome.


Smart and scalable solution


The solution can be scaled to make effective use of video at any level. One system allows you to manage up to 57,600 video streams. The system transmits video data to a single cloud-based management system that can analyze and control more than 1000 GV-VMS systems. On the other hand, in the client's data center or in the cloud, big data storage is available due to the use of servers based on Intel processors.
The Smart Video Management Solution also integrates with other systems, such as fire safety or access control systems, expanding their overall functionality.
By combining deep learning and integration with other hardware and software, cities can use solutions like the GeoVision solution to modernize their video surveillance systems. Deep learning improves automatic response, and integration improves operational efficiency. Scalability in turn means that urban growth will not affect the capabilities of video systems.
When it comes to urban use of video systems, traditional video surveillance solutions satisfy basic needs. But as soon as the project grows to the level of an entire city, then in just one day you can get thousands of hours of recordings. In this case, you will need too many people and time to detect the necessary events. The GeoVision solution makes it easier for operators to find the right objects or events.