Researchers from Honda, the University of Michigan, and Indiana University developed a deep learning system that can predict the trajectories of vehicles at road intersections.
“Safe driving requires not just accurately identifying and locating nearby objects, but also predicting their future locations and actions so that there is enough time to avoid collisions,” the researchers stated in their paper. “Our work is the first to address the problem of future vehicle localization under egocentric view and challenging driving scenarios such as intersections,” the team explained.
Using NVIDIA Tesla P100 GPUs, with the cuDNN-accelerated Keras and TensorFlow deep learning frameworks, the researchers trained a recurrent neural network on a new dataset of first-person videos collected from a variety of scenarios at intersections. The dataset includes 230 videos taken in over 2,400 vehicles.
“We use the gated recurrent unit GRU as the basic RNN cell. Compared to long short term memory (LSTM), GRU has fewer parameters, which makes it faster without affecting performance,” the researchers said.
Although their method generally performs well, uneven road surfaces and pedestrians blocking the view can throw it off. The team says future work could avoid this type of error by better modeling the entire traffic scene as well as the relation between traffic participants.
“Future work includes incorporating evidence from scene context, traffic signs/signals, depth data, and other vehicle-environment interactions. Social relationships such as vehicle-to-vehicle and vehicle-to-pedestrian interactions could also be considered,” the researchers stated.
The work was published on ArXiv this week.