[Paper Review] An Evaluation of Trajectory Prediction Approaches and Notes on the TrajNet Benchmark

Stefan Becker, Ronny Hug+2 more|arXiv (Cornell University)|May 19, 2018

Video Surveillance and Tracking Methods32 references63 citations

TL;DR

The paper evaluates various deep neural networks for predicting pedestrian trajectories on the TrajNet benchmark and identifies a simple RNN-Encoder-MLP (RED-predictor) as a competitive baseline without scene cues.

ABSTRACT

In recent years, there is a shift from modeling the tracking problem based on Bayesian formulation towards using deep neural networks. Towards this end, in this paper the effectiveness of various deep neural networks for predicting future pedestrian paths are evaluated. The analyzed deep networks solely rely, like in the traditional approaches, on observed tracklets without human-human interaction information. The evaluation is done on the publicly available TrajNet benchmark dataset, which builds up a repository of considerable and popular datasets for trajectory-based activity forecasting. We show that a Recurrent-Encoder with a Dense layer stacked on top, referred to as RED-predictor, is able to achieve sophisticated results compared to elaborated models in such scenarios. Further, we investigate failure cases and give explanations for observed phenomena and give some recommendations for overcoming demonstrated shortcomings.

Motivation & Objective

• Assess the effectiveness of deep neural networks for single-trajectory pedestrian path prediction using TrajNet datasets.
• Analyze dataset properties and noise to establish robust baselines.
• Propose simple preprocessing and model adjustments to improve prediction without human interaction cues.
• Identify failure modes and provide recommendations to overcome limitations.

Proposed method

• Compare simple baselines (MLP, RNN variants, TCN) and more complex Seq2Seq architectures on world-plane pedestrian data.
• Use offsets/velocities as inputs and outputs to stabilize learning and reflect motion dynamics.
• Train with 8 observed positions to predict next 12 positions using MSE loss in TensorFlow with Adam optimizer.
• Evaluate using ADE and FDE as common trajectory forecasting metrics.
• Analyze dataset noise via spline-based smoothing to estimate ground-truth noise.
• Develop RED-predictor: RNN-Encoder with a dense MLP on top for final predictions.

Experimental results

Research questions

• RQ1 What is the maximum achievable prediction accuracy on TrajNet World H-H data using baseline deep networks without human interaction cues?
• RQ2 How do simple preprocessing choices (offset-based inputs, standardization) affect prediction performance compared to more complex interaction-based models?
• RQ3 Can a simple RNN-Encoder-MLP reach performance comparable to elaborate models that use social or scene context?

Key findings

• The RED-predictor (RNN-Encoder with a dense MLP) achieves competitive results versus more elaborate models.
• Across architectures, predictions are often close to a linear interpolation baseline, with no clear single best model.
• Predicting full paths (not stepwise) helps mitigate error accumulation in sequential predictions.
• Offsets/velocities as inputs improve stability and performance over raw positions.
• RNN-Encoder-MLP often outperforms many baselines, but multiple architectures yield similar performance.
• Scene context and human-human interactions can further improve accuracy beyond single-trajectory models.