Researchers at the University of Michigan Develop Neural Networks

[Erica Zhu Feilong Jiangli]

Chongqing Feilong Jiangli collects the latest information for you, so that you can get the most cutting-edge technical information. Researchers at University of Michigan, by focusing on human gait, body symmetry and foot position, use more precise technology than now to teach autopilot to identify and predict pedestrian actions. Data collected by vehicles through cameras, lidars and GPS allow researchers to capture video clips of human action and reproduce them in three-dimensional computer simulations. On this basis, scientists have created a "biomechanically inspired recurrent neural network" to classify human actions.

According to the researchers, they can use the neural network to predict the movement and future position of one or more pedestrians 50 yards away from the car. If a vehicle wants to have the necessary predictive ability, the neural network needs to study the details of human movements in depth: the rhythm of human gait, the mirror image of symmetrical limbs and the way in which the position of feet affects human stability when walking.

Many machine learning algorithms that upgrade the level of automatic driving to the present level involve two-dimensional images. If a computer is shown millions of pictures with parking signs, it will eventually be able to recognize them in real time. By using a few seconds of video clips, the University of Michigan system can study the first half of the video and make predictions, then confirm the predictions with the second half of the video.

The research results show that this new system improves the ability of the autopilot to predict the most probable situation in the future. Finally, the system can enhance the safety of the self driving vehicle.