Prediction – Benchmark Models

This page brings together Benchmark Models contributed by the community, following a similar format. Prediction projects generally include main.py (main function), README (dataset-download and environment-setup guidance), Data (raw data), Map (map data), Api (official API interfaces), model (model files), utils (utility functions), result (program results), preprocess.py (preprocessing), train.py (training), test.py (testing), and visualize.py (result visualization).

LaneGCN

  • Description: Vehicle trajectory prediction using the Argoverse dataset
  • Overview: Combines road information with a lane graph convolutional network to predict multimodal vehicle trajectories in different traffic conditions and the confidence of each possible trajectory.
  • Link:https://github.com/TOPSlearningcenter/lane_1

PLM

  • Description: Non-motorized road-user trajectory prediction using the Xianxia Road–Jianhe Road dataset
  • Overview: Cyclists are modeled as individuals with heterogeneous interaction strategies. Preference learning captures these strategies, and combinations of heterogeneous preferences describe interactions to guide trajectory prediction.
  • Link：https://github.com/TOPSlearningcenter/Prediction-XXJH-PLM_1

MemoNet

  • Description: MemoNet, an instance-based（instance–based)method that predicts an agent’s motion intent by finding similar scenarios in the training data.
  • Overview: MemoNet has a memory bank that explicitly stores representative training instances to simulate the prefrontal cortex, together with a trainable memory addressing mechanism（memory addresser), which adaptively searches the memory bank for instances similar to the current context, simulating basal-ganglia function. During prediction, MemoNet retrieves relevant instances through the memory addressor. It first predicts the target location and then completes the full trajectory from that predicted location.