[Agtech Rising] GINT Attaches an ‘AI Kit’ to Existing Farm Machinery to Prevent Agricultural Work Accidents

According to the Ministry of the Interior and Safety’s Disaster Yearbook, 4,708 agricultural machinery accidents occurred over the four years from 2020 to 2023. By year, there were 1,269 cases in 2020, 1,076 in 2021, 1,384 in 2022, and 979 in 2023, with around 1,000 accidents recurring every year.

The accidents are concentrated among older drivers. According to a study by the National Institute of Agricultural Sciences on the status of agricultural machinery traffic accidents, 37% of such accidents involved drivers aged 70 or older, followed by drivers in their 60s at 27.6% and those in their 50s at 17.6%. Most causes were human factors such as failure to keep a proper lookout, misjudgment, operational mistakes, and drinking. The analysis is that as people age, slower judgment and response lead to accidents.

The aging of rural communities is worsening. In Statistics Korea’s (now the National Data Office) Agriculture, Forestry and Fisheries Survey, among 973,707 farm households nationwide in 2024, farms whose operators were aged 70 or older accounted for 50.8%, more than half. Including those in their 60s (33.0%), the figure reaches 83.8%. With the farm population shrinking and the workload handled by each person increasing, this is why critics say it is difficult to reduce accidents with safety management that relies only on human attention.

Agricultural autonomous-driving company GINT has stepped forward to solve this problem with AI. GINT formed a consortium with Agmo and Dverse and is carrying out the project “Agricultural Work Accident Zero! Development and Demonstration of an AI Agricultural Machinery Safe Operation Service” as an AX (Artificial Intelligence Transformation) project of the National IT Industry Promotion Agency (NIPA). The project runs from January 2025 (year one) to December 2026 (year two). The ultimate goal is to “AX-ify” agricultural machinery by attaching an “AI integrated solution kit” to tractors, sprayers, and other equipment that farms already own, without requiring them to buy new machinery.

Warnings when attention wanders; detours or stops when obstacles appear

The kit focuses on two types of risk. One is driver inattention, and the other is unexpected situations encountered during work. A driver-seat camera detects drowsiness, mobile phone use, smoking, and whether a seat belt is being worn. An external camera determines whether there are sudden obstacles or terrain that agricultural machinery cannot enter, warns the driver, and tells them how to respond. GINT Principal Engineer Jeong Hee-jong explained, “It is a model that prevents accidents by making users as aware as possible of risks stemming from inattention or insensitivity to danger.”

GINT said obstacle response has already been implemented to a level that does not require human intervention. For example, if an obstacle appears ahead while a ridge has already been made on the left, the kit makes its own judgment and detours to the right. If it cannot detour, it stops the machine.

The learning model was initially built using public data from AI Hub, and its performance was improved through field demonstrations. According to GINT, it has exceeded all first-year targets: 75% obstacle recognition accuracy, route regeneration within 10 seconds, and an 85% detection rate for risky driver behavior. Jeong said, “It varies by item, but we have secured reliability of more than 90%.” The second-generation kit currently on sale does not have this recognition function. The AI model will be included starting with the new product to be released next year. Even after launch, the company plans to update the learning model in the cloud and send it down to terminals in real time.

Alarm at a 30-degree tilt; forced stop if it keeps moving for another 5 seconds

Jeong explained that when rollover accidents occur, people are often working alone and are discovered too late, resulting in many fatalities. The kit sounds an alarm when the terminal tilts by about 30 degrees. If the machine continues moving for five seconds after the alarm sounds, it is forcibly stopped. After that, if the angle changes sharply, such as tilting by more than 10 degrees within one second, an emergency message is sent to predesignated contacts. Jeong said, “Going forward, through policy proposals, we are also reviewing ways to link it with emergency reporting systems such as 119 or 112.”

50 million won for a new-vehicle option, 15 million to 20 million won for the kit

The reason for choosing the method of attaching the kit to existing agricultural machinery is cost. Jeong said, “By modifying a tractor only minimally, it can deliver autonomous-driving performance comparable to that of an expensive tractor,” adding, “For new agricultural machinery, the autonomous-driving option costs about 50 million won more, but the kit can be installed for 15 million to 20 million won, saving 30 million to 35 million won in initial adoption costs.” It can also be installed on older models without electronic control units. The consortium has set a target price at around 25% of commercially available high-end autonomous agricultural machinery. Based on its experience installing devices on more than 300 types of domestic and overseas agricultural machinery, it also believes the kit can be applied to more than 90% of agricultural machinery in Korea.

While carrying out this project, GINT also signed agreements with agricultural machinery manufacturers such as LS Mtron Co., Ltd., Hada Co., Ltd., and Asia Technology Co., Ltd. The agreements state that if the kit proves performance above a certain level, the manufacturers will also adopt the technology.

From resistance to acceptance… demonstrations at eight rental centers

Since September, demonstrations have been underway at eight agricultural machinery rental centers, with one machine at each site: seven tractors and one sprayer (SS machine). The reason for choosing rental centers was data. Jeong explained, “Individual farms operate tractors for less than 600 hours a year, but rental centers operate them three to four times more than that,” adding, “We chose rental centers to collect as much data as possible in the early stages.”

The field response was not positive from the start. Jeong said farmers who had handled agricultural machinery for a long time tended to have a habit of driving while distracted and felt great inconvenience whenever warnings appeared, adding, “At the beginning of the project, some even told us to take it off.” There was also strong resistance to machines stopping regardless of the user’s will, so for some farms, the stop function was removed and only the detour function was included.

Now, such complaints have subsided. Jeong said, “There were comments during the initial familiarization process after installation, but currently there are no farms filing claims.” However, cases of misrecognition continue to be reported, such as when sunlight reflects off vinyl and is mistakenly recognized as a puddle, so the company is addressing them through retraining. Citing Tesla’s Full Self-Driving (FSD) as an example, he said, “There may be people who have never tried it, but no one uses it only once,” adding, “Once users start to feel the convenience of even reducing one finger movement, trust in the product will build.”

Data retention and responsibility remain challenges

The project also leaves challenges. Camera footage generates about 20 GB per day based on eight hours of operation. The kit terminal does not store data and sends only necessary data directly to the server. Because of personal information issues, in consultation with the demonstration farms, data used for learning is deleted once learning is complete. Jeong said, “What remains from an AX project is video data, and it is regrettable that we cannot retain it, making it difficult to use when developing other products.”

It has also not yet been determined who will be responsible if the AI misses a danger and an accident occurs. Jeong pointed out, “Clear rules are needed on whether the manufacturer or the user will be responsible.” GINT explained that once a system is in place to determine the responsible party, it will be able to apply functions more actively in line with the control capabilities it has already secured. As the budget for the agricultural machinery rental business has been transferred to local governments starting this year, the conditions for dissemination will also vary depending on local government finances. GINT is consulting with local governments while also making proposals to the Ministry of Agriculture, Food and Rural Affairs so the technology can lead to subsidized projects after certification.

The expected effect is limited to accidents within farmland. Because the kit’s autonomous response function is designed on the premise of work on farmland, accidents during road driving are excluded. Jeong said, “When an accident is predicted, the final judgment the machine makes is to stop,” adding, “If it detects nine out of 10 cases, then unless the user takes separate action, it will stop in all nine cases, so if the recognition rate is 90%, the prevention rate can also be seen as 90%.” He added, “As data accumulates and updates continue, it will rise to 91% and 92%,” and “If we guarantee performance to the extent that users trust the kit, I believe 100% can also be achieved.”

However, this prevention rate applies only to situations in which the kit can recognize danger. Machines without the stop function will not stop even if they detect danger, and accidents not captured by the camera, such as a body being caught in a rotating part, are not included in the calculation.

Regarding future plans, GINT Senior Researcher Noh Dong-won said, “The project ends in December this year, so we will sequentially end the pilot service starting in late November, and we plan to transfer ownership of the outputs developed through this project to the ministries and agencies that conducted the project.” He added, “We plan to commercialize the AI functions we developed by applying them to next-generation products in a way that farmers can use without resistance.”