What the Constant Velocity Model Can Teach Us About Pedestrian Motion Prediction

Abstract

Pedestrian motion prediction is a fundamental task for autonomous robots and vehicles to operate safely. In recent years many complex approaches based on neural networks have been proposed to address this problem. In this work we show that -surprisingly -a simple Constant Velocity Model can outperform even state-of-the-art neural models. This indicates that either neural networks are not able to make use of the additional information they are provided with, or that this information is not as relevant as commonly believed. Therefore, we analyze how neural networks process their input and how it impacts their predictions. Our analysis reveals pitfalls in training neural networks for pedestrian motion prediction and clarifies false assumptions about the problem itself. In particular, neural networks implicitly learn environmental priors that negatively impact their generalization capability, the motion history of pedestrians is irrelevant and interactions are too complex to predict. Our work shows how neural networks for pedestrian motion prediction can be thoroughly evaluated and our results indicate which research directions for neural motion prediction are promising in future.

Crowd Density Forecasting by Modeling Patch-Based Dynamics

“…Despite the extensive work done to leverage deep recurrent models for trajectory forecasting, a simple approach that makes future trajectory predictions based on the velocity is considered a powerful approach. Following [43], we implemented a simple baseline that used the velocity at the last two frames of input sequences to linearly extrapolate future trajectories. • Social LSTM (S-LSTM) [15] is one of the most popular baselines used in many trajectory forecasting papers.…”

Human motion trajectory prediction: a survey

“…Several recently introduced metrics follow a sampling approach to evaluate a probability distribution. The minimum average displacement error (mADE) metric (Rhinehart et al, 2019; Schöller et al, 2019; Tang and Salakhutdinov, 2019; Thiede and Brahma, 2019; Walker et al, 2016), as well as variety loss, oracle, minimum over N , best-of- N , top n% , or minimum mean squared distance (minMSD), computes Euclidean distance between the ground-truth position of the agent s t * at time t and the closest (or the n % closest) of the K samples from the predicted probability distribution: min k ∥ s t * − s t k ∥ . Similarly, minimum final displacement error (mFDE) evaluates only the distribution at the prediction horizon T .…”

“…Several authors report a separate accuracy measurement for the more challenging (e.g., non-linear or anomalous) part of the test set (Fernando et al, 2018; Huynh and Alaghband, 2019; Kooij et al, 2019), or evaluate the model’s performance on different classes of behavior, e.g., walking or stopping (Saleh et al, 2018b). Analysis of generalization, overfitting, and input utilization by a neural network, presented by Schöller et al (2019), makes a good case for robustness evaluation.…”

A Location-Velocity-Temporal Attention LSTM Model for Pedestrian Trajectory Prediction

“…To achieve real-time performance, their method uses parallelizable convolutional layers and incorporates no social or scene information. Another example is the method of Schöller et al [26] where the authors revisit and use the simple constant velocity model to predict the relative displacements between consecutive location points. In our previous work [27], apart from the velocity information computed directly from the input trajectories, our joint Location Velocity Attention LSTM based network (LVA) also requires no neighbourhood or scene information.…”