LSTM Networks for Human Trajectory Prediction in Simulated Crowded Scenes with Embedded Group and Obstacle Information

Abstract

The analysis of crowded scenes is one of the most interesting research areas in visual surveillance due to the wide variety of factors that affect such an analysis: environment structure and details, occlusions, obstacles, and others. More traditional methods such as Kalman filters relied on one-step forecasting but solutions are emerging that use recurrent neu ral networks to learn spatial-temporal dependencies of moving agents in crowded scenes. Recent works have used Recurrent Neural Networks (RNNs) and group sociology to improve the performance of the trajectory prediction task by segmenting users based on the measure of the coherency of their motion. Pedestrians are clustered into a single unit if they show similar or coherent motion patterns. This clustering better models social dependencies between and among agents in a visual scene, thus improving the performance of the prediction task. We build on this work, extending it to simulated crowds and explore embedding additional information about the environment and social psychology to further improve the results of the Long Short-Term Memory (LSTM) networks on the prediction task.