Digital Media Professionals Inc. (DMP) conducted a demonstration experiment of an abnormal behavior detection system utilizing LLMs at Grand Green Osaka, achieving an 83.5% detection accuracy for nuisance skateboarding and zero incidents of damage for the month. The background behind this initiative includes facility management challenges such as rising patrol costs in vast public spaces and the limitations of human responses.

Background of the Demonstration Experiment and Challenges in Public Spaces

In large-scale public spaces such as municipal parks, commercial facilities, and construction sites, there is a serious challenge of having to ensure safety across vast areas with limited management personnel. In particular, nuisance skateboarding and reckless riding in skateboard-prohibited areas have become a major headache for facility managers, as they cause risks of collision accidents with pedestrians and damage to pavement and street fixtures. Conventional countermeasures have centered on manual responses such as installing physical barriers and regular patrols by security guards; however, these incur substantial costs and have clear limitations in terms of prompt responsiveness across expansive grounds and the continuity of nighttime monitoring.

To address these challenges, the implementation of cutting-edge technologies through public-private partnerships has been promoted at Grand Front Osaka (Kita-ku, Osaka City), a pioneering project under the “Osaka Smart City Strategy” advanced by Osaka Prefecture and Osaka City. As part of the Smart City Implementation Support Project (FY2025) overseen by the Umekita 2nd District Smart City Formation Council, Digital Media Professional Inc. (DMP) and Nishio Rent All Co., Ltd. collaborated to plan a demonstration experiment for “NoSkate Ai,” an abnormal behavior detection system utilizing generative AI and large language models (LLMs). Utilizing the framework of the “MIDORI Partner System” operated by the Umekita MMO general incorporated association, full-scale technical verification set in an actual urban space has thus begun.

The Impact of Reaching 83.5% Detection Accuracy and Achieving Zero Nuisance Behavior

The demonstration experiment was conducted for approximately four months, from January to April 2026, in a section of Grand Green Osaka. The system works by transmitting footage from installed surveillance cameras to the cloud, where a cloud-based LLM evaluates behavior in real time. At the start of the experiment in January 2026, detection accuracy remained at just 41.8%, with frequent false positives mistakenly identifying shadows of people falling on the pavement or shadows on the roadway as actual humans. Especially during nighttime hours, a total of 20 false positives occurred over a 63-day tracking period (approximately 0.32 cases per day), making it an urgent priority to improve accuracy for practical implementation.

To address this, DMP analyzed massive amounts of video data collected in actual operating environments and improved the detection logic multiple times. Furthermore, it newly introduced a double-check system (two-stage determination) in which generative AI re-evaluates the results of primary detection. Through this tuning, detection accuracy as of the end of April 2026 achieved a dramatic improvement from 41.8% at the initial stage of the demonstration to 83.5%. Regarding the number of false detections at night, it decreased to just 3 cases (approximately 0.14 cases per day) over the 22 days following the logic improvements, keeping the occurrence frequency to less than half. The transition of accuracy improvements and the number of detections in the demonstration experiment is as follows.

Furthermore, starting February 27, 2026, we began operating a feature that automatically plays warning audio from local speakers upon detection. As a result, the number of skateboarding detections decreased from 71 in February 2026 to 55 in March, and dropped to 0 cases for the month in April 2026. By combining automatic warnings with automated email notifications to administrators containing still images and videos, it has been demonstrated that extremely high immediate deterrent effects against nuisance behavior can be achieved without relying on security guards for constant patrols.

Next-Generation Abnormal Behavior Detection Enabled by LLMs and Vision AI

“Vision-LLM Insight”: Decoding the Meaning and Context of Actions

The foundational technology behind this demonstration experiment is “Vision-LLM Insight,” an action recognition AI platform that DMP has been offering in full-scale since September 2025. With conventional vision AI and general object detection technologies, while it was possible to identify the presence of a “person” or a “skateboard” in video footage, it was difficult to distinguish subtle behavioral differences, such as whether the person was walking while carrying the skateboard by hand or riding it with their feet in a restricted area. Consequently, there were frequent cases where false alerts were triggered simply for carrying the items in.

“Vision-LLM Insight” adopts a hybrid structure in which DMP’s proprietary vision AI narrows down candidates for analysis through high-speed object detection, and then an LLM (Large Language Model) deeply interprets the surrounding context and the meaning of actions within the video. This makes it possible to correctly understand specific, complex action patterns—such as “riding and skating on a skateboard”—from the context and make highly accurate determinations. Furthermore, this platform supports flexible scenario design tailored to each customer’s individual challenges, achieving optimal processing that combines edge and cloud computing.

A Two-Stage Evaluation Mechanism for Generative AI That Overcomes Shadows and Lighting Variations

In actual urban environments, the irradiation angle of sunlight and the luminous intensity of streetlights change constantly depending on the season, time of day, and weather. “Complex shadows cast on the road surface” and “fluctuations in nighttime illumination”—which are unpredictable in a development environment or a single experimental facility—have been the biggest factors causing false positives in video analysis AI. To address this challenge, DMP has built a two-stage determination (double-check) algorithm utilizing generative AI.

As a first step, vision AI and LLMs quickly perform a primary extraction of potential abnormal behaviors from wide-area camera footage. Next, as a second step, generative AI thoroughly re-evaluates the extracted video frames, assessing whether “it is a false positive caused by shadows or lighting” and whether “the subject is actually performing a sliding motion.” By incorporating this dual verification process, we have successfully and dramatically reduced false alerts even in complex outdoor environments. Iterative tuning based on operational data collected across diverse times of day, weather conditions, and crowd volumes has led to high stability in real-world environments. The operational concept of the detection system in the field is as follows.

Operational Benefits and Future Outlook Toward Societal Implementation

Reducing implementation costs and ease of deployment through the use of existing cameras

One of the major strengths of the abnormal behavior detection system demonstrated this time is that it can leverage existing infrastructure assets as they are. There is no need to newly install expensive dedicated AI cameras; the system operates simply by transferring video streams from general-purpose IP cameras or security cameras already installed within the park or facility to the cloud. This makes it possible to rapidly deploy the latest LLM-based action recognition solution while minimizing initial capital expenditure.

Furthermore, rental availability for this system began in October 2025 through Nishio Rent All, accommodating flexible operational formats such as short-term use limited to event durations or temporary deployment at setup sites. When an anomaly is detected, facility managers simultaneously receive an email with attached still images and video as a warning sound goes off from the on-site speaker. This enables them to instantly grasp accurate on-site conditions even from a remote location and take necessary responses with minimal effort.

Expansion into Multifaceted Action Scenarios and the Future of Smart Cities

Based on the extensive real-world operational data and systematic insights into false-positive patterns obtained from the demonstration experiment at Grand Green Osaka, DMP and Nishio Rent All are further expanding the functional capabilities of “Vision-LLM Insight” and “NoSkate Ai.” Moving forward, they plan to go beyond the detection of nuisance skateboard riding and drive development to support a wide range of behavioral scenarios, including unauthorized entry into restricted areas, abandoned hazardous materials, suspicious loitering, and pedestrians falling or collapsing due to sudden illness.

This technology is expected to be applied in all fields where crime prevention and safety management are required, such as municipalities nationwide, managers of parks, roads, and rivers, large-scale commercial facilities, construction sites, and logistics warehouses. In particular, collaboration with Nishio Rent All, which has a broad network of bases and a customer base nationwide, will accelerate its social implementation as smart city social infrastructure. Initiatives to enhance the safety and comfort of urban spaces using AI technology while taking personal privacy into consideration are drawing significant attention as a new disaster prevention and crime prevention model in the era of Physical AI.

Reference Page

• [DMP News Release] https://www.dmprof.com/ja/news/pr-20250616-ai-skateboarder-detection-system/
• [DMP News Release] https://www.dmprof.com/ja/news/pr-20250908-vision-llm-insight/
• [Nishio Rent All Press Release] https://www.nishio-rent.co.jp/news/?m=Detail&id=530

[#DMP #LLM #AICamera #AnomalousBehaviorDetection #GrandGreenOsaka #SmartCity #NishioRentAll]