1. Program background
Drowning prevention has become a "compulsory course" in safety education for teenagers and a shared responsibility for safety management among all sectors of society. Especially when the weather gradually turns hotter after entering summer, this is also a dangerous, prone and high-incidence period for drowning accidents. Traditional prevention and management methods are through daily publicity and lectures on the dangers of drowning and manual patrol management, which have many disadvantages:
1) Lack of effective safety warning facilities: When people approach the danger range, there is a lack of danger reminders and monitoring and supervision.
2) It is difficult to control personnel: the water area is open, the waterfront is long, and the activity area is large, making it difficult to control personnel behavior.
3) Backward management methods: The existing equipment is simple and primitive and cannot accurately identify, locate and handle emergencies efficiently.
In order to strengthen early warning of drowning risks, improve various safety protection facilities and strengthen inspections, so as to detect dangers in a timely manner, we can use the intelligent means of "anti-drowning algorithm" to carry out intelligent supervision and risk early warning of waters to ensure the safety of people's lives. Today I would like to introduce to you the anti-drowning AI solution based on TSINGSEE Qingxi video technology and artificial intelligence technology.
2. Plan Overview
TSINGSEE Qingxi Video AI algorithm platform deploys 45 AI algorithm models. In anti-drowning supervision scenarios, it can be implemented through algorithms for people to break into dangerous waters. This algorithm supports the identification of people entering the warning area on the premise of delineating anti-drowning areas (which can be divided into dangerous areas, warning areas, and safe areas), and at the same time, reporting whether they are barefoot, clothed, and age characteristics, etc. information.
Relying on the video footage collected by the surveillance cameras deployed at the water site, the TSINGSEE Qingxi Video AI algorithm platform can conduct real-time analysis of the accessed video streams, monitor in real time whether anyone breaks into the warning zone, and can trigger alarms in time to remind managers If handled in a timely manner, the on-site voice device can also be linked to broadcast the eviction.
The principle of the anti-drowning AI algorithm is mainly an AI algorithm that detects whether there are people in the warning area (ROI). The input is an image or video frame, and the ROI is set (ROI is a closed polygonal area). The algorithm automatically calculates whether a pedestrian is within the ROI. The main basis for determination is whether the center of the pedestrian's detection frame is within the ROI. It is mainly used in scenes where there is a small flow of people, pedestrians’ limbs are clear, and the obstruction is not serious; the recommended number of people ranges from 3 to 10 people.
3. Solution value
1) AI + ordinary camera intelligent anti-drowning system, all-weather real-time monitoring and early warning
Deploy anti-drowning video surveillance systems in hazardous water scenes such as reservoirs and rivers. The TSINGSEE video visual monitoring AI algorithm is trained on "thousands of drowning photos" and can conduct 24-hour uninterrupted detection of managed waters and accurately identify Whether there are people next to dangerous waters, once someone enters the designated dangerous area, the system will automatically send out a voice alarm to persuade people to leave, and automatically push a reminder message to the management staff. The management staff can give remote warnings and arrive in time. On-site, avoid dangerous incidents to the greatest extent, ensure the safety of personnel, and effectively improve the management and control of dangerous waters.
2) Achieve precise monitoring and management to reduce the rate of drowning accidents
On the basis of traditional "human defense + physical defense", an intelligent technological defense line of "technical defense" can be added. Based on the video capabilities of the security monitoring system EasyCVR platform, it can realize 24-hour video monitoring and risk warning, real-time monitoring of dangerous areas such as rivers, lakes, and reservoirs, and improve the management efficiency of drowning prevention. TSINGSEE's video anti-drowning AI solution uses edge AI vision technology to identify behaviors such as people approaching dangerous waters, conducts intelligent analysis and video aggregation and fusion of monitored video images, provides fast real-time feedback, and is highly intelligent.
4. Scenarios and suggestions
1) Camera setup
In all application scenarios, try to unify the height and angle of the camera installation. The algorithm cannot adapt to all camera angles and heights. Because when pedestrians are at different heights and angles, the posture and size changes of pedestrians vary greatly.
The ideal scene is a slightly overlooking scene, not a large overlooking scene (special optimization is required). The pedestrian aspect ratio is 1:1.5 ~ 1:2.5. Slightly looking down can avoid some occlusions between pedestrians, and at the same time ensure that the changes in pedestrian scale proportions will not be drastic. When the camera looks too straight, the drawn distant area is easily invaded by nearby pedestrians, and the regional effect is prone to false alarms, failing to achieve the desired effect. The camera setup should be as unified as possible to ensure that a set of algorithms can be scene-compatible on video images from each camera to achieve better results.
2) Camera internal parameters
The focal length of the camera should be controlled to ensure that pedestrians are no less than 50*100 pixels in a 720P image.
3) Image quality
The video bit rate is higher, the captured video frames or images have higher resolution, and the limbs of pedestrians are clearer. High-resolution images are very necessary for data annotation and model prediction. The light is relatively good and pedestrians can be better distinguished from the background.