The University of Zurich's breakthrough in vision-based autonomous systems shows that vision-based autonomous systems can indeed outperform humans through the use of advanced neural networks and machine learning algorithms.
Most autopilot solutions usually use a combination of radar and cameras to make perception more accurate. However, there are a number of drawbacks to this approach. In their work, engineers have found that adding additional sensors, such as radar, to the camera will affect the final analysis rate and results, and lengthy programs** will also make the driving system not "smart" enough.
In addition, there are more and more vehicles equipped with lidar, which will interfere with each other and affect the detection accuracyThe observed value of the radar will change from time to time, which will bring a certain errorIn addition to that,When the information transmitted by the camera and the lidar conflicts, who should the car machine choose to believe?"If the information analyzed by the radar with the camera contradicts each otherIt will be difficult to choose an intelligent driving system. It's better to choose only one and do it to the extreme.
Musk insists on getting rid of inertial thinking and starting from first principles: roads are designed for biological neural networks and eyes, people do not have organs similar to radar, and only eyes can still recognize traffic on the road, then cars can do the same. Andrej, former senior director of AI at Tesla, said: ".It is hoped that a neural network similar to the visual cortex of animals can be built to simulate the process of information input and output in the brain.
Just like light entering the retina, we wanted to simulate this process with a camera. It can be seen that Tesla is committed to developing a pure vision self-driving system, which can achieve higher driving performance than humans by simulating the structure and function of the human eye and brain.
In fact, back in 2020, Tesla announced that it would abandon radar and fully embrace a camera-based self-driving system. At that time, there were still many doubts in the industry, but with the further improvement of algorithms and computing power, pure vision autonomous driving solutions are rapidly achieving breakthroughs, and the research of the University of Zurich only further confirms this.
Studies have shown that by mimicking the structure of the retina of the human eye and the visual cortex of the brain, a huge convolutional neural network can not only meet or even exceed the performance level of humans in traditional visual tasks such as image classification and object detectionand can convert the image input into driving decision-making output end-to-end to complete the entire autonomous driving perception and decision-making process.
In other words, this is an artificial intelligence solution that highly mimics a biological vision system from input to output. The input terminal simulates eye image acquisition, the middle is used to extract and understand features through a convolutional neural network similar to the visual cortex structure, and finally the output terminal directly generates driving decisions, without the need for traditional multi-sensor fusion or rule engine.
The study also shows that this pure vision system can carry out fast and accurate traffic environment perception, including target detection, tracking, motion estimation and other functions, and the speed and quality of perception even exceed the human level. In driving simulations and road tests, it has demonstrated strong and stable autonomous driving capabilities.
What's more, this whole process of understanding the traffic environment and making decisions on driving is completely based on efficient neural network deep learning algorithms, which replaces the inefficient and cumbersome manual feature engineering and rule set design of the past, and is fully replicable, scalable, and sustainable. This means that with more computing power and datasets, the drivability of such systems will continue to increase exponentially.
Although there is still some distance from true commercialization and application on busy streets, pure vision autonomous driving systems are clearly the most promising solution. It is highly biomimetic, and at the same time, it is also an efficient deep learning solution, which is bound to become the mainstream and trend in the development of this field. Tesla is taking advantage of the trend and is fully promoting this strategy, and I believe that it will be the first to succeed.
The research at the University of Zurich is a key breakthrough technology that has burst out in this process. It shows surprising capabilities in both system design and practical results, confirming the enormous potential of pure vision systems to surpass humans in autonomous driving. It is believed that in the near future, such a system will change the face of transportation and mobility for the benefit of human society.