LAformer: Trajectory Prediction for Autonomous Driving with Lane-Aware Scene Constraints

This repository contains the official implementation of LAformer: Trajectory Prediction for Autonomous Driving with Lane-Aware Scene Constraints

Quick Start

Requires:

• Python ≥ 3.6
• PyTorch ≥ 1.6

1) Install Packages

 pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
 pip3 install -r requirements.txt

2) Install Argoverse API

The latest version of Argoverse requires Python ≥ 3.7

If using Python 3.6, you can install Argoverse v1.0

https://github.com/argoai/argoverse-api

3) Install Nuscenes

 pip3 install nuscenes-devkit

1. Download the nuScenes dataset. For this project we just need the following.

   • Metadata for the Trainval split (v1.0)
   • Map expansion pack (v1.3)

2. Organize the nuScenes root directory as follows

   └── nuScenes/
   ├── maps/
   |   ├── basemaps/
   |   ├── expansion/
   |   ├── prediction/
   |   ├── 36092f0b03a857c6a3403e25b4b7aab3.png
   |   ├── 37819e65e09e5547b8a3ceaefba56bb2.png
   |   ├── 53992ee3023e5494b90c316c183be829.png
   |   └── 93406b464a165eaba6d9de76ca09f5da.png
   └── v1.0-trainval
       ├── attribute.json
       ├── calibrated_sensor.json
       ...
       └── visibility.json         

Usages Argoverse 1

1) Train (First Stage)

Suppose the training data of Argoverse motion forecasting is at ./train/data/.

OUTPUT_DIR=checkpoints/models.laformer.1; \
GPU_NUM=8; \
python src/train.py --future_frame_num 30   --do_train --data_dir ./train/data/ \
--output_dir ${OUTPUT_DIR} --topk 2 --hidden_size 128 --train_all  --vector_size 32 \
--train_batch_size 128 --use_map  --num_train_epochs 10 --lane_loss_weight 10  --core_num 32 \
--subdivide_length 5    --use_centerline --distributed_training ${GPU_NUM} \
--other_params  semantic_lane direction step_lane_score enhance_global_graph  point_level-4-3

Add --reuse_temp_file to skip re-listing the map file for the second time running or use --temp_file_dir ${path} to specify the pre-processed data path.

1) Train (Second Stage Motion Refinement)

Suppose the training data of Argoverse motion forecasting is at ./train/data/.

OUTPUT_DIR=checkpoints/models.laformer.s2.1; \
GPU_NUM=8; \
MODEL_PATH=checkpoints/models.laformer.1/model_save/model.10.bin; \
python src/train.py --future_frame_num 30   --do_train --data_dir ./train/data/ \
--output_dir ${OUTPUT_DIR}   --hidden_size 128   --topk 2 \
--vector_size 32 --train_batch_size 128 --use_map  --num_train_epochs 9 \
--lane_loss_weight 10  --core_num 32 --subdivide_length 5  --learning_rate 0.0003\
--train_all --use_centerline --distributed_training ${GPU_NUM}  \
--other_params  semantic_lane direction step_lane_score enhance_global_graph \
point_level-4-3  stage-two-train_recover=${MODEL_PATH} stage-two-epoch=9 stage_two

2) Evaluate

Suppose the validation data of Argoverse motion forecasting is at ./val/data/.

Example:

OUTPUT_DIR=checkpoints/models.laformer.s2.1; \
GPU_NUM=1; \
python src/eval.py  --future_frame_num 30 --eval_batch_size 128\
--output_dir ${OUTPUT_DIR} --hidden_size 128 --train_batch_size 128 \
--lane_loss_weight 10 --topk 2 --use_map  --vector_size 32 --model_recover_path  9 \
--core_num 16 --use_centerline --distributed_training ${GPU_NUM}\
--subdivide_length 5 --other_params semantic_lane direction step_lane_score enhance_global_graph \
point_level-4-3  stage-two-train_recover=${MODEL_PATH} stage-two-epoch=9 stage_two \
  --do_eval  --data_dir_for_val ./val/data/

Usages NuScenes

1. Run the following script to extract pre-processed data. This speeds up training significantly.

   python src/datascripts/dataloader_nuscenes.py --DATAROOT path/to/nuScenes/root/directory --STOREDIR path/to/directory/with/preprocessed/data 

   1) Train (First Stage)

   Suppose the processed training data of NuScenes is at checkpoints/models.laformer.nuscenes.1/temp_file.

   OUTPUT_DIR=checkpoints/models.laformer.nuscenes.1; \
   GPU_NUM=8; \
   python src/train.py --do_train --future_frame_num 12 --output_dir ${OUTPUT_DIR} \
   --topk 2 --hidden_size 64 --train_batch_size 32 --num_train_epochs 50 \
   --lane_loss_weight 0.9 --distributed_training ${GPU_NUM} --reuse_temp_file \
   --other_params  semantic_lane direction step_lane_score enhance_global_graph \
   point_level-4-3 nuscenes nuscenes_mode_num=5

   Add --reuse_temp_file or use --temp_file_dir ${path} to specify the pre-processed data path.

1) Train (Second Stage Motion Refinement)

Suppose the processed training data of NuScenes is at checkpoints/models.laformer.nuscenes.1/temp_file.

OUTPUT_DIR=checkpoints/models.laformer.nuscenes.1; \
GPU_NUM=8; \
MODEL_PATH=checkpoints/models.laformer.1/model_save/model.50.bin; \
python src/train.py --future_frame_num 12   --do_train --output_dir ${OUTPUT_DIR}\
--hidden_size 64   --topk 2 --train_batch_size 32  --num_train_epochs 50 \
--lane_loss_weight 0.9   --reuse_temp_file --distributed_training ${GPU_NUM}  \
--other_params  semantic_lane direction step_lane_score enhance_global_graph \
point_level-4-3  stage-two-train_recover=${MODEL_PATH} stage-two-epoch=50 stage_two \
nuscenes nuscenes_mode_num=5

To train a model with 10 modes, you can set nuscenes_mode_num=10, --topk 4, and --num_train_epochs 100, the other settings is the same as aforementioned command.

2) Evaluate

Suppose the processed test data of NuScenes is at checkpoints/models.laformer.nuscenes.1/temp_file.

Example:

OUTPUT_DIR=checkpoints/models.laformer.nuscenes.1; \
GPU_NUM=1; \
MODEL_PATH=checkpoints/models.laformer.1/model_save/model.50.bin; \
python src/eval.py  --future_frame_num 12 --eval_batch_size 128 \
  --output_dir ${OUTPUT_DIR} --hidden_size 64 --train_batch_size 32 \
  --lane_loss_weight 0.9 --topk 2 --reuse_temp_file --model_recover_path  50 \
 --distributed_training ${GPU_NUM} --other_params step_lane_score  stage_two \
    semantic_lane direction enhance_global_graph subdivide new laneGCN \
  point_level-4-3 stage-two-train_recover=${MODEL_PATH} stage-two-epoch=50 \
nuscenes nuscenes_mode_num=5 --do_eval

Pretrained Models

We provide the pretrained LAformer in checkpoints/. You can evaluate the pretrained models using the aforementioned evaluation command.

Results

Quantitative Results

For this repository, the expected performance on Argoverse 1.1 validation set is:

Qualitative Results

Citation

If you found this repository useful, please consider citing our work:

@inproceedings{liu2024laformer,
  title={Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints},
  author={Liu, Mengmeng and Cheng, Hao and Chen, Lin and Broszio, Hellward and Li, Jiangtao and Zhao, Runjiang and Sester, Monika and Yang, Michael Ying},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2039--2049},
  year={2024}
}

License

This repository is licensed under Apache 2.0.

Awesome Related Repos

• https://github.com/Tsinghua-MARS-Lab/DenseTNT
• https://github.com/ZikangZhou/HiVT