According to foreign media reports on February 14, scientists at the University of Michigan are using human gait, body symmetry and foot displacement to teach autopilot cars to identify and predict pedestrian actions with higher accuracy than the prior art.

It is reported that the data collected by the camera through the camera, lidar and global positioning system (GPS) enables researchers to capture video clips of human activities and then reproduce them in three-dimensional (3D) computer simulations. On this basis, scientists have created a “revolutionary neural network inspired by biomechanics” to classify human movements.
According to the researchers, they use the recurrent neural network to predict the movement of one or more pedestrians about 50 yards from the car and its future location, which is equivalent to the size of a city intersection. In order to have the necessary predictive power, the car needs to use the cyclic neural network to study the details of human movements, including the rhythm (periodicity) of human gait, the mirror symmetry of the limbs, and the influence of the position of the foot on the stability of the human body during walking. .
Ram Vasudevan, assistant professor of mechanical engineering at the University of Michigan, said: "The previous research in this field usually focused on still images and not on how people move in three dimensions. However, if these cars are to operate and interact in the real world, we There is a need to ensure that the predictions of the location of the pedestrian are inconsistent with the next step of the vehicle. The movements of the pedestrians and the places they look at can tell you the level of their attention and tell you what they are going to do next."
Most machine learning algorithms that take autonomous driving technology to the current level involve two-dimensional images—that is, still photos. If a computer displays millions of parking sign photos, it will eventually be able to identify the stop sign in real time in the real world. However, by using a video clip that runs for a few seconds, the system can study the first half of the video clip for prediction and then use the second half to verify the accuracy.
In the end, the research results show that this new system enhances the ability of driverless vehicles to predict the most likely occurrence in the future, which is conducive to improving the safety of self-driving cars.