Reborn in 2010, the AI ​​in my mind is too powerful.

Bain Capital was very accommodating; even before their investment in the United States was finalized, they only received a letter of intent before transferring Uniview in China to Chen Yiran.

The equity change of Uniview Technologies has been filed with the Hangzhou Binjiang District Administration for Industry and Commerce. Wanguoyuan holds 100% of the equity of Uniview Technologies, and the registered address has been changed from a temporary office in Binjiang District to the sign of Wanguoyuan Hangzhou Branch.

Moreover, they only received 5 million in the end, which was actually lower than their acquisition price. This clearly showed their intention to leave a good impression on Chen Yiran, facilitating future cooperation.

It must be said that Bain Capital is a breath of fresh air among various investment forces.

After parting ways with Bain & Company, Chen Yiran originally considered simply buying H3C.

However, HP showed no intention of selling, so we had to give up and wait for another opportunity.

Now, Chen Yiran has already achieved all of her initial goals in the country.

He was called to Hangzhou by the longan project team from Wanguoyuan. Yin Qi and Tang Wenbin had already done a lot of visual training work for Wanguoyuan that year, and their experience was coming in handy this time.

In a meeting room, Yin Qi and Tang Wenbin from the Longan project team, Wang Tao from DJI, and Zhang Guopeng from Uniview sat below, watching Chen Yiran on top of them make his demands.

He first looked at Yin Qi.

"The first thing is to re-identify pedestrians across cameras. A person walks in from the east gate of the park and is captured by a bullet camera at the east gate. Five minutes later, he appears at the entrance of the cafeteria and is captured by a PTZ camera outside the cafeteria."

The system needs to know that this is the same person and doesn't need to re-identify. Your current three models—face, license plate, and intrusion detection—can only perform single-point recognition.

Cross-camera tracking requires training a separate re-identification model. OpenCV and previous spatiotemporal constraint rules can't achieve this level of accuracy; we need to rebuild the framework using DeepFlow. I have the technical expertise for this; I'll send it to you later. You can then perform on-site optimization and testing at Shougang Park.

Normally, such a breakthrough, starting from scratch in the industry, would be quite shocking if Chen Yiran suddenly said that he had technical reserves.

However, only Wang Tao and Zhang Guopeng seemed somewhat surprised, while Yin Qi and Tang Wenbin calmly nodded.

They've been at Wanguoyuan for a year now and have witnessed countless miracles. Chen Yiran will reply to their emails with questions on a regular basis, although the response time is slow, it always sends chills down their spines, making them feel like they've wasted over a decade of education.

Many things, they just try to ask, and then Chen Yiran just throws a complete version for them to follow. The questions are just things he wrote down casually before.

"The second thing is to identify abnormal behavior on the front end."

Existing intrusion detection methods can only determine if someone has entered an area they shouldn't be in, but cannot determine what that person is doing.

The algorithm labeling logic is different for three behaviors: one person squats down, one person runs, and one person puts their bag on the ground and walks away.

Upgrade Longyan's current intrusion detection model to include six behavioral categories: running, crouching, loitering, throwing objects, climbing, and falling. Mr. Zhang, your NVR motherboard originally had 128MB of RAM; running these six behavioral categories definitely requires expansion. Calculate the time yourself, align it with Yinqi's, and report back to me.

Zhang Pengguo thought for a moment and said, "Running a deep learning model that classifies six types of behavior, even if your framework is very clean, it still requires at least several hundred megabytes of memory, which current embedded chips simply cannot handle."

Chen Yiran nodded: "I made some arrangements beforehand, and Wanguoyuan has some technical reserves. We will customize a special NVR motherboard and configure it with better chips."

"The third thing is the automated patrolling by drones."

We plan to build a drone hangar at each of the four corners of the Shougang Park. Each hangar will have a built-in automatic charging base. After the drone lands, it will lock and automatically charge, then stand by. Flight paths will not need to be manually set; the system will automatically plan patrol routes based on feedback from the fixed cameras at Longyan.

For example, this afternoon, three cameras on the west side of the park malfunctioned and went offline. The system not only needs to remind people to check manually, but also automatically dispatches an aircraft from the nearest hangar to replace the aerial coverage.

Wang Tao felt a bit of pressure. Although these things didn't sound complicated, they actually far exceeded the current technical standards for drones. However, looking at the three people next to him who looked like it was a matter of course, he felt that he couldn't lose face, so he reluctantly nodded and agreed.

"The fourth thing is that I don't just want to use facial recognition."

After the system went online, everyone who entered the camera's field of view in Shougang Park would be recorded by the re-identification model, which would record a code after fusing gait features and clothing features.

"A is the same person, B is the same person, C is the same person—the system knows where they go through multi-dimensional technical comparisons. I also have the technical reserves for this. Although it's something I use at home, the framework is fine; it just needs more data to feed it. Longyan and Uniview, you need to cooperate on this aspect."

After Chen Yiran finished writing on the blackboard in a long and detailed manner, he turned around and asked, "What are your thoughts? Speak up now, and tell me if you have any difficulties."

The four men stared at each other for a while, until Zhang Guopeng, the oldest, raised his hand and asked, "Boss, from what I've seen so far, our job seems to be the easiest."

Let me ask on behalf of those three young people: for example, this drone hangar, isn't that a bit too difficult? Although I don't know much about it, it's hard to imagine that modern drones can achieve that level of precision.

Furthermore, this multi-dimensional technical comparison, which includes not only gait but also, according to you, height, hairstyle, and various other factors, does the current algorithm truly support that?

Before Chen Yiran could say anything, Yin Qi spoke first: "I don't know about drones, but the boss said the algorithm is possible, so it definitely is. Our Longan project team has made a lot of achievements in the past year, most of which are used in wafer fab production."

Although the specific things being identified differ, the core framework is largely the same; the DeepFlow framework can achieve this, and it's incredibly powerful.

Zhang Guopeng nodded upon hearing this: "Okay, what about the drone issue?"

Chen Yiran didn't speak, but looked at Wang Tao, gesturing for him to say a few words.

Zhang Guopeng, Yin Qi, and Tang Wenbin also turned to look at Wang Tao.

Wang Tao instantly felt immense pressure, but as a technological pioneer, he thought for a moment and said, "Our current accuracy is definitely not high enough, and using signals for positioning is not very realistic."

However, since the boss said that we have technical reserves in vision, I think we can mainly use vision calibration? This kind of approach is not uncommon and is very common in college robot competitions. Although the recognition distance cannot be compared with actual applications, the idea is the same.

Furthermore, the University of Michigan just published a paper this year about a technology called AprilTag, which is applicable to robot vision localization, camera calibration, and augmented reality systems. We can place an AprilTag or a checkerboard pattern on the top of the hangar.

When the drone returns to its nest, it first uses GPS for positioning. When it lands in the airspace, it switches to bottom camera recognition of the marker. The Longyan visual model calculates the position and distortion angle of the marker in the picture in real time, and outputs the XYZ offset of the drone relative to the marker to the flight controller for the final alignment correction in the last few meters.

This should be feasible, but laser radar and real-time dynamic carrier phase differential technology are not feasible due to the drone's payload capacity. Also, regarding the battery, what are our battery life and charging requirements? These would need to be specifically customized.

Chen Yiran nodded: "No problem, there's no need to use AprilTag. I have the technical skills for that, you guys can match it later."

Regarding batteries, I've already acquired an American company, and I'll be acquiring a Chinese one later, as our battery supplier. We can communicate with them and make technical requests if any problems arise. Okay, is everything alright now?

Wang Tao nodded, hesitated for a moment, and then asked a question that had been puzzling him for a long time: "Boss, can we have a bigger target? Although technically speaking, we should have no problem with this job."

However, from a business perspective, I feel that fully commercializing what we've ultimately achieved would be quite difficult. Shouldn't we conduct some market research first?

Chen Yiran waved his hand: "Since you asked, I'll tell you about the positioning of your companies. In my mind, the priority for your projects is technical implementation! Making money is just secondary. Everything is based on my technical requirements, understand? I don't need you to consider the market."

"I have plenty of money! All I need is for you to run at breakneck speed along the route I've laid out for you!"

After the meeting, the four of them went back to work, each with their own thoughts.

From a technical perspective, they are very happy to have a boss like Chen Yiran who doesn't care about the market but only focuses on technology.

On the other hand, Chen Yiran's technology is completely one-man rule, which Wang Tao and Zhang Pengguo are not used to. Of course, Yin Qi and the other two have been criticized a lot and have come to think that this is normal. A genius like the boss should naturally be in control of the whole situation.

Chen Yiran ignored them. He had been working non-stop for two weeks and was planning to take a break and go see his old friends.