【2021 SUSE 國壹（伍）】Review 16th National College Student intelligent car racing competition of iFLYTEK

Write in front

Time flies. It’s the end of the year in the twinkling of an eye. The most impressive thing this year is the smart car competition in the summer vacation. It’s time to review some problems and solutions. It’s also a response to someof your problems and make a simple sharing and exchange.
-----------------------------------------------------------------------------------------------------2021.12.31 HYJ YiBin

What do I do

I’m just a humble child watching movies. I was called to identify myself in this competition. Hahaha, so I don’t know how to drive and hardware related knowledge. (don’t scold me)

Some understanding of the rules of the competition

The biggest feature of this track is that you don’t have to make your own car! Directly use the car developed by iFLYTEK!

Personally, I think this is fair, at least in terms of hardware. It also tests the use of specific technology, engineering technology and strategies. Especially ROS platform and target detection.

One more thing, it’s nice not to make your own car, because I’ve known other groups… Really, the car condition, train of thought and degree of realization are really different. Maybe this is the charm of other tracks, hahaha.

Drive car?

Again, I can’t drive. I’m just a movie watcher. Driving is the responsibility of other big men in our group. It’s very powerful.

Object Detection（Here comes mine. Let’s get to the point）

Let’s start with a few things:
① The car’s own camera is a wide-angle camera, so there is imaging bending (it will also exist to a certain extent after timely correction, after all, it has physical characteristics, so there is no way)

② Camera default mirror effect!!! Be sure to OpenCV image conversion when opening the camera, otherwise you will make an error in identifying the QR code! (the official does not directly explain this situation. You need to try it yourself. Pay attention to the default image problem)

③ The detection object is really small and fuzzy. It is not what I imagined and in reality. It is a simple target detection task. (I also made a whine that can detect glasses and hair in my life)

Therefore, we used some strange methods to complete it.

Legend method？

Go directly to the target detection method used in the final national game. Personally, I think it is very unique.

Q1: why use the python version of Yolo target detection?
A1: as the saying goes, tensorflow in industry and pytorch in academia. This sentence is not false, but also the existing situation. It can be seen that tensorflow has relatively high requirements for the hardware itself. The smart car is just a nano, so in order to ensure efficiency and accuracy, pytoch, which is more suitable for ARM architecture, is used.

Q2: why use yolov3?
A2: Tiny！！！ Relatively small, suitable for nano with low computing power and low power.

Q3: why not use yolov5?
A3: the reason why V5 is not used is that CUDA versions do not match, which does not play the power of V5. It seems that other teams can only use the first generation of V5, so they chose the simpler V3 architecture. (after all, FPN is less than V5.)

Q4: about parking and taking photos?
A4: all possible locations for detection shall be photographed for detection. Or, for global time estimation, use time At the sleep interval, the arrival position of the vehicle is estimated at the specified time for photographing.

Q5: legendary detection strategy?
A5: just read it. Turn complexity into simplicity and return to nature.
Address: https://pan.baidu.com/s/1RSeY81qI2ZvosIjahgFsNA
Password: 0315

Unfortunately, I didn’t repeat the steps. Personally, I think it’s too difficult because the objects themselves have little difference, the pixels are low, and the amount of computation will increase, so I didn’t do it. Here I pay tribute to Comrade Zuo Qingyu of Nanjing University of Aeronautics and Astronautics for his detection strategy. He used to remove duplication. I admire it very much. Unfortunately, he lost the national game in the end.

Something wanna say

Thank you for meeting and experience. This competition is the application of knowledge, the competition of engineering technology, the cooperation between teams and interdisciplinary, and the opportunity to communicate and compete with other schools. Have experienced countless failures, experiments, staying up late, laughing, success, tiredness and self blame. Fortunately, we came to the end. I saw the on-the-spot failure of other teams and lost their qualification to enter the national competition. I also saw the vehicle problem in the final finals, resulting in the loss of competition opportunities and regret to leave.

When I was preparing for the national game in August, I went home for examination and recuperation due to physical reasons. I didn’t fight with my teammates at school. I had to solve it remotely. Time was tight. During this period, both psychological pressure and physical condition once made me want to give up the game and get a minimum living allowance reward. Finally, I would like to thank brother Zhonghua for his repeated urging and team sense of honor. I stayed up three days and solved the part of target detection I was responsible for.

Final

2021 is about to pass. Thank the people around me. I wish you a happy new year and a bright future.

                                                                     感謝相遇，感謝經歷，止於至善，盡善盡美

-------------------------------------------------------------------------------------------------------------2021.12.31 23:28 HYJ ENZO