Demonstration of Low-Load Data Collection and Communication Control Technology for Image Recognition Processing of Multiple High-Resolution Camera Feeds on the IOWN APN

Contributing to the reduction of on-site workload through accurate wide-area monitoring of wild birds and animals

News Highlights:

Background

In recent years, the introduction of remote monitoring systems, where camera images are transmitted over a network to a server and processed by AI for image recognition, has been advancing as a countermeasure to the shortage and aging of workers. Examples include wild animal monitoring to detect animals such as bears and wild boars, vehicle monitoring to measure traffic volume and manage parking lot entry and exit, and intrusion detection to identify suspicious individuals or unauthorized access to restricted areas.

However, due to limitations in network and server load capacity, increasing the number of cameras used for data collection or improving image resolution is difficult, resulting in localized monitoring only. Specifically, in current wild animal monitoring, surveillance is limited to specific points such as checking the status of traps, and there is no continuous or wide-area monitoring that enables tracking of species and population numbers. As a result, on-site tasks such as regular patrols of forest areas still rely on manual labor.

To reduce the burden of such labor-intensive fieldwork, there is a growing need to achieve wide-area monitoring without human intervention by deploying numerous high-resolution cameras and analyzing the images they capture (Figure 1).

Among the constraints of deploying cameras, those related to the network side are expected to be resolved with the advancement of next-generation communication technologies such as IOWN APN and 6G. On the other hand, on the server side, there is still no established method to mitigate the processing load required to receive large volumes of data. Therefore, there is a pressing need to establish a technology that enables low-load collection of large volumes of data.

Figure 1 Current Status of Wildlife Monitoring and Future Vision of the Communication Control Technology

Demonstration experiment

In this demonstration experiment, NTT Central Training Center (Chofu, Tokyo) and eXeField Akiba (Chiyoda, Tokyo) were connected via the IOWN APN, and an experimental system was constructed in which images captured by multiple 4K cameras installed at the NTT Central Training Center were transmitted to a server installed at eXeField Akiba for AI-based image recognition. In addition, three use cases were simulated for image recognition: wild animal monitoring, vehicle monitoring, and intrusion detection (Figure 2).

This system adopts a configuration in which multiple cameras installed in a single building or area share the same network line and transmit image data to a server located remotely. To simulate an environment with a large number of installed cameras, the system combines actual terminals connected to physical cameras with virtual terminals that emulate cameras.

In the experiment, server load, throughput, and image recognition accuracy during image data collection were measured.

Figure 2 Operation of the Demonstration Experiment System and the Communication Control Technology

Key points of the technology

One of the factors causing increased processing load during the reception of large volumes of data is the memory copy process from the Operating System (OS) to the application. As a solution to this issue, the application of RDMA², a technology that directly writes data from the memory of one server to the memory of another server, is considered. RDMA is characterized by high speed and low load; however, it operates on the premise of lossless communication paths, requiring the construction of mechanisms to suppress packet loss within the network.

Conventionally, in data center networks where RDMA has been used, packet loss has been prevented by coordinating flow control functions across all communication devices installed nearby. However, in the wide-area networks targeted for this data collection, where various communication devices are deployed extensively over a large area, it is difficult to operate the same coordinated functions. Consequently, packet loss cannot be suppressed, and directly using RDMA results in a significant degradation of communication performance.

Therefore, a low-load data collection communication control technology (hereinafter referred to as proposed communication control technology) was established to suppress packet loss and secure communication performance by controlling the start and end timing of RDMA communications generated from multiple terminals via a controller. In proposed communication control technology, terminals request the controller to start communication triggered by the occurrence of data to be collected. The controller then controls the start and end timing of RDMA communications based on each terminal's request so that communications do not collide, thereby suppressing packet loss.

Furthermore, considering the characteristics of RDMA such as its high speed and the diversity of delay requirements depending on the collected data, the communication start and end are performed immediately only for data requiring prompt analysis. For data that can tolerate longer analysis times, communication start and end are carried out by aggregating a certain amount of data to increase system throughput. This behavior minimizes the overhead of timing control and further enhances communication performance.

Thus, by applying communication control suitable for data collection to RDMA, which is both high speed and low load but requires lossless communication paths, it becomes possible to collect a large volume of data from many sources with low load while maintaining communication performance.

Results of the demonstration experiment

When collecting data using conventional RDMA, although server load reduction is possible, packet loss occurs due to line sharing among many terminals, causing a significant degradation in communication performance. On the other hand, when collecting data using proposed communication control technology, packet loss during RDMA use can be suppressed, and throughput was confirmed to improve by approximately five times. Additionally, compared to data collection using TCP³, it was confirmed that server load for reception processing can be reduced to as low as 1/1000 while maintaining comparable communication performance (Figure 3).

The effects of applying proposed communication control technology include an increase in the number of cameras supported and improvement in resolution. In the case of TCP, the number of supported cameras is limited due to server load constraints; however, by using proposed communication control technology, server load becomes negligibly small, allowing the number of supported cameras to be increased up to the maximum available network capacity. Therefore, it is expected that the number of cameras supported can be expanded by approximately ten times⁴.

In addition to realizing wide-area monitoring by increasing the number of cameras, high-accuracy detection is also expected through image recognition using high-resolution camera images such as 4K. This enables detection of subjects even when they appear small due to factors such as the animal's small size or distance from the camera. In this experimental system, the detection rate for small subjects was approximately 60% when filmed in Full HD, whereas it improved to approximately 80% when filmed in 4K, confirming an accuracy improvement of about 20%.

By enabling continuous wide-area monitoring through the collection of many high-resolution camera image data, the application scope of remote monitoring is expanded, leading to a reduction in on-site work burden. In particular, in wild animal monitoring, this technology is expected to extend to advanced analysis such as the population, habitat range, and movement paths of animals, thereby reducing the workload of regular patrol surveys and contributing to solving issues such as worker shortage and aging.

Figure 3 Server Load and Average Throughput

Role of each company

Future Developments

This time, we demonstrated that numerous high-resolution camera image data can be collected with low load while maintaining communication performance using the low-load data collection communication control technology.

Going forward, by leveraging the IOWN APN and this technology, we will continue technology development to explore applications not only in wild animal monitoring but also in further remote monitoring fields, contributing to solving social issues such as the fundamental reduction of on-site work.

[Glossary]

¹IOWN APN (All-Photonics Network)
An innovative network based on photonics technology, with its architecture being openly developed by the IOWN Global Forum. By expanding the application scope of photonics (light)-based technology, it enables transmission with low power consumption, high quality and large capacity, and low latency, which are difficult to achieve with current electronics (electron)-based technologies.
https://www.rd.ntt/e/iown/

²RDMA (Remote Direct Memory Access)
A technology that directly transfers data in memory over a network. It is used for short-distance data transfer in data centers and high-performance computing.

³TCP (Transmission Control Protocol)
A communication protocol widely used in IP (Internet Protocol) communications. It achieves highly reliable data transfer through control functions such as congestion control and retransmission control.

⁴Assuming CPU resources of four cores and a network bandwidth of 400 Gbps available for data collection communication.

⁵Press release August 30, 2024: Wild animal damage countermeasures in cooperation with local hunting associations
https://www.ntt-east.co.jp/yamagata/new/detail/pdf/20240830_01.pdf (Japanese)

About NTT

NTT contributes to a sustainable society through the power of innovation. We are a leading global technology company providing services to consumers and businesses as a mobile operator, infrastructure, networks, applications, and consulting provider. Our offerings include digital business consulting, managed application services, workplace and cloud solutions, data center and edge computing, all supported by our deep global industry expertise. We are over $90B in revenue and 340,000 employees, with $3B in annual R&D investments. Our operations span across 80+ countries and regions, allowing us to serve clients in over 190 of them. We serve over 75% of Fortune Global 100 companies, thousands of other enterprise and government clients and millions of consumers.