AI cameras at the edge are expected to expand their market, focusing on public institutions and the security market. While the manpower and capabilities of monitoring agents are limited, experts expect that AI will be effective in the field as it can mechanically learn and infer set elements.

The 2023 CCTV Integrated Control Conference was held at KINTEX in Ilsan on the 29th. Hosted by the Ministry of the Interior and Safety, this event was attended by officials from Hanwha Vision, Korea Internet & Security Agency, and Seoul Metropolitan Government, who gave presentations on the use and introduction of intelligent CCTV and AI technology.

The announcements that day were attended by Hanwha Vision, which announced the technological performance trends of AI cameras applied to intelligent CCTV, and Seoul City Hall, which announced suggestions for operating intelligent CCTVs in local governments.

■ Edge-end AI camera increases license plate recognition efficiency
“The most efficient way to apply AI technology to CCTV is to recognize vehicle license plates. However, if AI images are analyzed on the server side, it may be difficult to implement at 4K level.”

Hong Jeong-bin, Vice President of Hanwha Vision, explained that they encountered such a difficult problem during the development of intelligent CCTV.

He said that the biggest problem when transmitting 4K video during the development stage of vehicle recognition was bandwidth, and mentioned that 4K video transmission, which requires a bandwidth of up to 12 Mbps, consumes twice the storage capacity compared to 2K video, and the burden on the center's GPU server for video analysis is four times as much.

Hanwha Vision solved this problem with an AI camera at the edge. With the 'Best Shot' method, it independently obtains the best shot of a passing vehicle's license plate and the vehicle, matches it with the license plate of the same vehicle, crops it from JPEG, and sends only the image to the center, thereby alleviating bandwidth and the burden on the center.

In order to secure the next crop image from Best Shot, it is necessary to secure at least 120 pixels, and the width that can be obtained with 120 pixels is approximately 7.2 meters. This Best Shot method can be equipped with a facial recognition function if additional costs are invested.

Hanwha Vision already enables people search through facial recognition Best Shot, and when searching for specific elements of a person, such as 'woman wearing red', related results can be derived as Best Shot. This can be applied equally to license plate recognition, and it is expected that this will be able to assist in various public affairs such as △tracking criminal vehicles and △searching for missing persons.

■ Intelligent CCTV, all bad… “The field is constantly changing”
“Current intelligent CCTVs are all bad. This is because they do not understand distance and changes over time.”

Lim Dong-hyun, a senior official at Seoul City Hall, pointed out that intelligent CCTV is not omnipotent and that there are problems with misrecognition and distance measurement. Furthermore, the problem was that the actual CCTV installation environment was different from the development environment, so even if intelligent CCTV was introduced, the efficiency would be low.
CCTVs are installed in large numbers, but in order to increase the efficiency of intelligent CCTVs by 100%, it seems necessary to devise a plan that can respond to changes over time, such as seasonal and spatial changes.

△It is expected that intelligent CCTV will have limitations in performing its functions due to various on-site variables such as poor CCTV angle, objects blocking the camera, low image quality and color expression errors, night lighting issues, and focus restrictions of single-focus products.

Chief Lim predicted that the reliability of intelligent cameras will increase if the night lighting conditions are improved by applying smart poles and fixed cameras are placed in appropriate locations to monitor the same point from different angles.

Lim Ju-moo-gwan suggested, “The current market direction is in the process of transitioning to processing both collection and analysis within CCTV, thereby reducing the center’s computing power.” He added, “In addition, learning must come first to improve technology, and there needs to be a social discussion on whether learning and disclosure of results for this purpose violates the Personal Information Protection Act.”

Meanwhile, according to the Korea Internet & Security Agency (KISA), the number of CCTVs in public institutions is estimated to be 1.4 million as of 2021, and is increasing by 10% every year. If the private sector is included, the number is expected to reach 16 million. Faced with the challenge of having to monitor 20 times more screens per control agent than the standard, the number of KISA intelligent CCTV-related certifications has been on a steady increase since the first certification in 2017, reaching 82 cases over the past three years.