Technical informationposition:Baoxingwei > Technical information > 正文
'TIANYAN' Smart Tracking warns of crime to Spot 'suspicious behavior' in crowds
Edit:Baoxingwei Technology | Time:2023-02-05 09:00 | Number of views:316
"Thirty seconds left." In a final reminder, Tom Cruise's future cop scrambles toward a house where an angry man holds up scissors to his cheating wife, and police break down the door and hold him down.
"Confirmed it's Howard Marx." The police identified the man by scanning his iris and said, "Pursuant to a warrant from the Center for Precrime, you are under arrest for the imminent murder of your wife, which will actually take place today at 8:04 a.m."
"But I didn't do anything! I'm not going to do anything!" The man shouted, but the police had him under control.
In the sci-fi movie Minority Report, where and when a murder will happen can be predicted, and in the future, police can arrive at the scene ahead of time, effectively stopping the crime.
With the rapid development of facial recognition technology, the 'standard feature of the future' in science fiction movies is becoming possible. Recently, the Financial Times reported that Chinese companies are helping police develop artificial intelligence that can be used to spot "suspicious" patterns of behavior in crowds, identify and arrest suspects before crimes are committed.
Cool technology cameras recognize behavioral trajectrajectory
Yao Zhiqiang, director of strategic planning at Yuncong Technology Co., is busy. The Financial Times report catapulted the facial recognition company onto the list of most-searched searches.
"We are building a recognition system where a person goes, what they do, and where their movements and extent are automatically recorded, stored and catalogued." Yao paused to emphasize that this is not the prediction of crime as foreign media said, "but the implementation of smart technology to prevent, control and track criminal behavior."
A staff member demonstrates the face data combat platform of Yuncong Technology. On the platform, the staff selected an area, then imported a photo of a tester's face, selected a time frame, and clicked Retrieve. See, on the expanded regional plane map, the places where the tester appeared have been marked by red coordinate symbols. Click any place, and you can see all The Times when the tester appeared there and the manner of dress at that time.
In the function of dynamic layout control, when the tester passes the camera, the system screen will pop up information in real time, including personnel category, time and place. According to the staff, in the practical application, the trajectory of the key personnel controlled by the police will be recorded and tracked in real time. Once they appear near the key places, the system will give an early warning and push the information to the duty desk.
In Yao's opinion, this seemingly cool technology is to solve a sub-problem of image retrieval: "Given a monitored personal image, retrieve the personal image across devices".
However, the implementation of the technology is not as easy as it sounds. "Dynamic recognition is very difficult." Yao explained that the open scene, the face will not take the initiative to cooperate with the camera Angle, plus resolution, light and other factors, the previous facial recognition accuracy is not high, "before the Guangdong Provincial Public Security Department has done a test, from a foreign system recognition rate of only 5%, that is to say, search 20 times may be able to find once."
In order to understand the differences between certificate photos and real faces, Cloud science and technology team personnel specially built a detachable structured data acquisition array, using 91 high-speed cameras, each with a 5 degree Angle difference to collect image data under different photography and lighting angles. After more than half a year, science and technology personnel added Angle correction, 3D modeling and other modules on the basis of the original algorithm, and constantly adjusted the parameters of each module through deep learning. "Now the recognition rate of a million magnitude of completely outdoor dynamic recognition can reach 50%".
They also run into trouble when the target's face is aging and intentionally wearing a mask such as glasses or a hat, making face recognition more difficult. "In addition to face recognition, artificial intelligence also covers long-distance gait recognition, voice print recognition, hair recognition and other biometric recognition methods, can be judged together."
To meet the need for traceable retrieval across devices, technicians also need to think about how to store massive amounts of video information efficiently.
"In the way of artificial intelligence, to achieve video structure." Yao Zhiqiang, the volume of video data is getting larger and larger. The structure of video can set the attributes of key information and store it in a highly compressed way. Specifically, when the target person passes by the camera, artificial intelligence will automatically recognize their identity information, separate out the face photos, and their activity trajectory will be automatically recorded into information mode, and classified and stored, "greatly facilitate the subsequent information retrieval".
The future of development depends on artificial intelligence for policing
"It's not suspicious for someone to buy a kitchen knife, but if they then buy a hammer and a bag, their risk rating goes up." Yao stressed that the assessment of criminal behavior of key management personnel is mainly tracked through the big data of faces. Once the activity data is continuously abnormal and cumulative risk level, the police can detect and evaluate in time, "real-time warning, prevention."
Yuncong is not the only company moving into Crimeprediction. Artificial intelligence has brought about the development of a new generation of technologies such as cloud computing and deep learning. Face recognition companies such as Sensetime and MegVII, which are good at algorithms, have also chosen their own "track" to work carefully.
In May this year, during the "Cultural Expo" held in Shenzhen, SenseFace system provided by SenseTime supported real-time face capture and recognition in 1000+ channels of surveillance video. Within 5 days, face recognition was completed for 210,000 people, 241 people were warned by comparison, and 25 people were convicted of criminal charges.
"The camera can replace human eyes for intelligent tracking and recognition of the target, and through artificial neural networks and other algorithms, to achieve the trajectory analysis of the passenger flow in a specified area." Xu Li, CEO of SenseTime, developed the intelligent crowd behavior analysis system based on the deep algorithm, which can carry out real-time crowd monitoring and make intelligent early warning by counting the number of people in the scene, tracking the movement speed and direction of the crowd, and analyzing abnormal behaviors.
The development of artificial intelligence, so that the security of the image demand from "visible" "see clearly", evolution into "understand". In fact, since 2009, the Ministry of Public Security has made it clear that graphic investigation (video investigation) has become the fourth major investigation technique after technical investigation, criminal investigation and network investigation. In recent years, Guangdong public security is also vigorously promoting the construction of "video cloud + big data" platform with facial recognition as the core, and asking for police force from artificial intelligence.
The cameras distributed in the major bayonets of the city are playing an increasingly important role. According to public media reports, 574,000 cameras and 1,590 sets of high-definition road bayonet systems have been installed in Guangzhou, realizing full coverage of major roads, key areas and key places.
"For crowded areas, when the camera collects every portrait passing by, the system will automatically describe its gender, wear and clothing characteristics, forming the corresponding feature big data of the portrait." Hong Xiaolong, chief of the Science and Technology Information Department of Guangdong Provincial Public Security Department, said in an interview with the media that cameras located at high places can detect key control personnel hidden in the crowd in time, their appearance time and location. For those who fall into the category of arrest, the alarm information will be proactively pushed to the mobile terminals of the front-line police to ensure that the police on the road will deal with them first.
The push for AI policing is getting bigger. On July 20, The State Council issued the "New generation of artificial intelligence development Plan" officially released, Vice Minister of Science and Technology Li Meng made it clear in a briefing that "in the field of public security, if intelligent security system is used, it can give advance warning, remind managers through big data, and provide decision-making basis for managers".
Debate continues over whether big data algorithms are fair
The Washington Post has long described the scene: Charles Coleman gets out of his police car and surveys the trash-strewn street in front of him, which is crowded with homeless people. He had come to deal with a "crime that has not yet happened".
It was not a 911 call that summoned the Los Angeles Foothills Police officer, but a tip from the crime prediction app Predpol that a car theft or break-in was likely in the area that morning.
Predpol is a set of crime-prediction software developed by George Moller, a mathematics professor at Santa Clara University in California, and a team led by him. Based on an equation that predicts aftershocks, the software is able to predict which areas are about to be hit by analyzing data on the type, time and location of previous crimes.
The technology has expanded to other parts of Los Angeles and more than 60 police departments across the United States, making it the most popular forecasting system in the country. According to Police magazine, the use of "predictive policing" data has improved the effectiveness of law enforcement agencies, with the Atlanta Police Department using the technology in two of its districts and seeing a 10 percent drop in crime within 90 days.
Predpol is not alone. In Tennessee, the Highway Patrol uses IBMSPSSModeler to predict where accidents will occur. In Milan, Italy, KeyCrime has been used for more than a decade to predict where robberies will occur based on historical data. SolutionHouse, a software solution developed in Cape Town, South Africa, predicts the likelihood of a crime.
In foreign countries, the application of crime prediction technology is developing beyond people's imagination. In September 2016, Stanford University released a periodic research report titled "Artificial Intelligence and Life in 2030". The report predicts that by 2030, predictive "police" will be in mass use and will eventually be relied upon by humans.
But the chorus of disapproval never stopped. Skeptics argue that there are thousands of possibilities for a person to commit a crime, and that what appears to be a fair algorithm is only judging the likelihood of a person committing a crime in the dimensions that its designers consider important.
Privacy and racial justice groups say there is little evidence that these technologies work, and they worry that the practice could unfairly focus law enforcement activities on communities of color by relying on racially biased patrol data.
The Stanford researchers also warned that there were still problems to be solved with the application of AI to justice. The first is data security, and the second is accuracy, "because algorithms are based on historical crime statistics to predict future crimes, there is a risk that past law enforcement patterns can be equated with the belief that certain people are predisposed to commit crimes."