Researchers created an algorithm that allows self-driving cars to predict lane-changes of the surrounding cars.
The system works by using a deep-learning technique called Long Short-Term Memories (LSTMs). With this technique, a neural network is trained to observe knowledge for a certain amount of time and propagate that knowledge to predict what will happen in the future. Although the technique is known to be unstable when getting long input sequences or predicting data far in the future, it works really well for limited time-scopes.
The input to the LSTM of the neural network consists of the relative positions and velocities from the closest six cars to the vehicle. The neural network takes this current state, looks at how the other vehicles behaved in the last 3 seconds. The network tries to predict whether the vehicle will move one lane left, one lane right, or will stay in the same lane. The LSTM outputs both a state of the LSTM and the predicted output of the LSTM. This output is passed through a neural network with two dense layers to end up at one of the three predictions.
The researchers varied the amount of data the neural network got between 1 second and 5 seconds and predicted up to 3 seconds in the future. The position and orientation of vehicles were detected at a distance up to 125 meters and collected with 25 measurements per second.
The authors of the paper wrote the paper as a way of predicting what other cars will do in the event of sensor-blackout: failure of the Lidar while still knowing where other cars will be. They trained and tested their algorithm on 500.000 lane-changes in an internal Toyota dataset. Although the most likely scenario on the highway is that every car stays in its own lane, their algorithm was able to slightly improve on this baseline prediction.
The authors published their paper on the arXiv.
