Meituan Unmanned Delivery Team Wins CVPR 2019 Obstacle Trajectory Prediction Challenge – Methodology Overview

Background : CVPR 2019, the premier computer‑vision conference, hosted a Trajectory Prediction Challenge under the Autonomous Driving Workshop. Meituan’s unmanned‑delivery and vision team won first place.

Challenge Description : Participants must predict the next 3 seconds of an obstacle’s trajectory based on its past 3 seconds of motion. Obstacles include pedestrians, bicycles, large vehicles, and small vehicles, sampled at 2 Hz.

Dataset : Training files contain 1 minute of multi‑modal (camera + radar) annotated data per road segment; each line records obstacle ID, class, position, size, and orientation. Test files contain 3 seconds of data, and the goal is to forecast the following 3 seconds.

Evaluation Metrics : Average Displacement Error (ADE) and Final Displacement Error (FDE) are computed per obstacle and then weighted‑summed across classes (WSADE).

Existing Methods : The problem is usually tackled by (1) independent prediction – using only each obstacle’s own history, and (2) interaction‑based prediction – leveraging cross‑obstacle information. Representative interaction models include Social GAN (encoder‑pooling‑decoder) and StarNet (star‑topology LSTM with a hub network).

Our Data Analysis : • Size information is irrelevant for position prediction, so it is ignored. • Orientation labels are noisy; we discard them. • Many trajectories are incomplete; we interpolate missing frames to obtain full 12‑frame sequences for training. • Data augmentation applies rotation, reversal, and Gaussian noise to each obstacle’s trajectory, since our method does not model inter‑obstacle interaction.

Model Architecture : We adopt a multi‑class independent prediction framework. For each obstacle class we build an LSTM encoder‑decoder. Between encoder and decoder we insert a Noise module that generates fixed‑dimensional Gaussian noise and concatenates it with the encoder’s final hidden state to initialise the decoder. The noise introduces stochasticity, allowing multiple plausible future trajectories.

Inference Rule : Among the generated trajectories we select the one whose direction best aligns with the historical direction.

Experimental Results : Training solely on the official data with the described augmentations yields a Weighted Sum ADE (WSADE) of 1.3425, outperforming StarNet (1.8626) and the ApolloScape baseline TrafficPredict (8.5881).

Conclusion : A multi‑class independent LSTM‑encoder‑decoder with noise augmentation achieved the top score in the CVPR 2019 Trajectory Prediction Challenge. Future work will explore interaction‑based and graph‑neural‑network approaches to further improve prediction accuracy.