Trajectory Prediction

The problem of trajectory prediction is essentially one of sequence prediction where given a sequence of input coordinates of a pedestrian, we must identify a pattern and use that to predict the future sequence of coordinates of the same pedestrian. This subsystem forms the crux of our entire system since the output of this subsystem is published to the vehicle.

We have made some good progress on this subsystem although we had to settle for a different approach than the one we had initially decided. The Social LSTM algorithm is the state-of-the-art in pedestrian trajectory prediction however it suffers from difficulties such as difficulty in debugging and dataset biases. Moreover the algorithm was unsuitable for our requirements in this semester since the we are focusing on a single pedestrian use case whereas the true value of the Social LSTM algorithm is found in cases with multiple pedestrians.

We hence shifted to a more deterministic approach of using polynomial regression. Based on an observation window of 12 frames (at a frame rate of 10 frames per second), we fit a polynomial curve to input values. Using this curve, we extrapolated for the desired 12 frames (at the same frame rate) and published this as our predicted trajectory. We analyzed this for various degrees of polynomials and observation lengths. Fig. x shows a comparison of some basic testing.

For the fall semester, we proceeded with a second degree polynomial fit. We tested this algorithm on live pedestrian data and were satisfied with the results that we were getting. The graphs below show the performance for some specific pedestrian trajectories.

Fig. x. Graphs of trajectory prediction for trajectories of different radii of curvature (r).
(a) r = ∞, (b) r = 3m, (c) r = 2m, (d) r = 1m, (e) r = 0m (right angle),