Autonomous driving is already being put into practical use as a car that automatically drives even if a human is not driving. It seems that each car manufacturer is conducting new empirical tests to improve behavioral safety. How is machine learning contributing to this autonomous driving technology? Voyage co-founder and CEO Oliver Cameron, who is developing autonomous driving technology, explains.
According to this, object detection technology has been mainly discussed among autonomous driving technologies for the past 10 years. Object detection technology is a technology that detects crosswalks and people walking by the side of a road, and if the accuracy of this technology is increased, the detection rate of people protruding in front of a car increases.
In the test of measuring object detection ability using ImageNet, which is a huge image data set, only 50% was achieved with cutting-edge technology around 2010, before machine learning became common. However, as of 2020, the accuracy is reaching 88%. It is making continuous progress
In 2012, when autonomous driving technology was still in its dawn, a research team of Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton applied a deep learning technology called AlexNet to object detection. Published a study.
Alexnet made a big impact on the autonomous driving industry with the highest accuracy at the time. As a result, around 2014, a number of autonomous driving technologies began to use deep learning. This trend continues until 2020. As a result of these advances, autonomous driving technology can safely detect important objects around it. Of course, Oliver Cameron says there is still work left to predict what the objects around him will do. If you make a good prediction, you can predict the movement of people or objects around you and behave appropriately, but if you make a mistake, it will cause an accident. Proper prediction may sound impossible, but humans are already acquiring information from thousands of people from their surroundings and making intuitive predictions.
An example of the current prediction problem of autonomous driving technology is the situation of making a right turn on a road. When making a right turn, it is necessary to predict the movement of vehicles in the counter lane and pedestrians at crosswalks and make a right turn. Therefore, turning right is a difficult problem even during autonomous driving. He describes the process when the current self-driving technology turns right. First, a recognition module such as a sensor detects a specific distance object and inputs the information into the prediction module. The second prediction module generates predictions of how each individual will move 5 seconds from the present, from the current and previous observations. Next, the individual movement question is substituted into all algorithms to calculate the safest action possible. Finally, put the safest action into action and re-evaluate the decision every 100 milliseconds.
He argues that autonomous driving performed in this way will have dangerous consequences. In particular, he points out that it can lead to the worst situation in a crowded city. Recently, there has been a movement to apply machine learning technology to prediction.
Although prediction of autonomous driving technology is not essential, it is explained that by applying machine learning technology to prediction, the decision-making ability of autonomous vehicles will be dramatically improved and passengers will be able to move safely. Of course, his company, Boise, is also planning to publish a study on predicting autonomous driving technology. Related information can be found here .