[Paper Review] MotionCNN: A Strong Baseline for Motion Prediction in Autonomous Driving

Stepan Konev, Kirill Brodt+1 more|arXiv (Cornell University)|Jun 5, 2022

Autonomous Vehicle Technology and Safety26 citations

TL;DR

MotionCNN presents a simple CNN-based baseline that outputs multiple future trajectory hypotheses with confidences, achieving competitive results on the Waymo Open Motion Dataset and ranking 3rd in the 2021 challenge.

ABSTRACT

To plan a safe and efficient route, an autonomous vehicle should anticipate future motions of other agents around it. Motion prediction is an extremely challenging task that recently gained significant attention within the research community. In this work, we present a simple and yet very strong baseline for multimodal motion prediction based purely on Convolutional Neural Networks. While being easy-to-implement, the proposed approach achieves competitive performance compared to the state-of-the-art methods and ranks 3rd on the 2021 Waymo Open Dataset Motion Prediction Challenge. Our source code is publicly available at GitHub

Motivation & Objective

• Motivate a simple, strong baseline for multimodal motion prediction without complex architectures.
• Leverage rasterized bird’s-eye-view representations and CNNs to predict multiple future trajectories with confidences.
• Demonstrate competitive performance against state-of-the-art methods on the Waymo Open Motion Dataset.

Proposed method

• Rasterize historical trajectories and map into multi-channel 224x224 images centered on the target agent.
• Use a CNN backbone (pretrained on ImageNet) to predict K trajectory hypotheses and their confidences via a mixture model.
• Model the future as a mixture of K Gaussians with fixed identity covariance; train with negative log-likelihood loss over the ground truth trajectory.
• During inference, output K=6 trajectory hypotheses and use their means as final predictions with associated confidences.
• Train with AdamW, cosine annealing with warm restarts, and early stopping; backbone variants (Xception71, ResNet18) evaluated.
• Raster caching to disk accelerates training by avoiding on-the-fly rasterization.

Experimental results

Research questions

• RQ1 Can a pure CNN-based raster image approach rival state-of-the-art motion prediction methods that use graphs or complex architectures?
• RQ2 Does a simple mixture-of-Gaussians loss with CNN regression provide meaningful multimodal trajectory predictions?
• RQ3 How does a lightweight backbone (e.g., ResNet18) compare to deeper backbones (e.g., Xception71) in this baseline?

Key findings

• The MotionCNN baseline ranks 3rd on the Waymo Open Motion Dataset motion prediction challenge using the mAP metric.
• It achieves competitive Min ADE and Min FDE scores, outperforming several more complex models on these metrics.
• With the Xception71 backbone, MotionCNN achieves mAP of 0.2136, Min ADE 0.7400, Min FDE 1.4936, Miss Rate 0.2091, Overlap Rate 0.1560 on the test set.
• The ResNet18 variant trains roughly three times faster but has lower performance than the deeper backbone.
• On the validation set, MotionCNN-Xception71 reaches mAP 0.2123, Min ADE 0.7383, Min FDE 1.4957, Miss Rate 0.2072, Overlap Rate 0.1576.
• The model outputs 6 trajectory hypotheses (K=6) corresponding to the evaluated prediction set size in the challenge.