PTP-DA

Official PyTorch code for the paper "Adaptive Pedestrian Trajectory Prediction via Target-Directed Data Augmentation".

Abstract: Pedestrian trajectory prediction is an important task for many applications such as autonomous driving and surveillance systems. Yet the prediction performance drops dramatically when applying a model trained on the source domain to a new target domain. Therefore, it is of great importance to adapt a predictor to a new domain. Previous works mainly focus on feature-level alignment to solve this problem. In contrast, we solve it from a new perspective of instance-level alignment. Specifically, we first point out one key factor of the domain gaps, i.e., trajectory angles, and then augment the source training data by target-directed orientation augmentation so that its distribution matches with that of the target data. In this way, the trajectory predictor trained on the aligned source data performs better on the target domain. Experiments on standard baselines show that our method improves the state of the art by a large margin.

Installation

Environment

• OS: Linux / RTX 2080Ti
• Python == 3.7.9
• PyTorch == 1.7.1+cu110

Dependencies

Install the dependencies from the requirements.txt:

pip install -r requirements.txt

Train

sh scripts/train.sh

Metrics

We use the ADE and FDE as the metrics. Here we give the whole results of 20 experiment groups.

ADE

Method	A2B	A2C	A2D	A2E	B2A	B2C	B2D	B2E	C2A	C2B	C2D	C2E	D2A	D2B	D2C	D2E	E2A	E2D	E2C	E2A	AVG
STGCNN	1.12	0.91	0.99	1.16	2.18	1.18	2.14	1.47	0.96	0.59	0.54	0.41	0.89	1.04	0.57	0.45	0.93	1.01	0.58	0.44	0.98
STGCNN+Ours	0.80	0.87	0.98	0.90	1.24	0.55	0.65	0.61	0.69	0.27	0.36	0.31	0.83	0.32	0.46	0.43	0.70	0.23	0.42	0.40	0.60
T-GNN	1.19	0.77	1.04	0.85	1.52	0.84	1.53	1.10	0.91	0.65	0.55	0.40	0.85	1.11	0.72	0.43	0.81	1.09	0.62	0.51	0.87
T-GNN+Ours	0.70	0.75	0.90	0.84	0.91	0.52	0.68	0.45	0.75	0.28	0.47	0.36	0.81	0.29	0.45	0.38	0.68	0.23	0.43	0.41	0.58

FDE

Method	A2B	A2C	A2D	A2E	B2A	B2C	B2D	B2E	C2A	C2B	C2D	C2E	D2A	D2B	D2C	D2E	E2A	E2D	E2C	E2A	AVG
STGCNN	1.49	1.07	1.34	1.49	3.21	1.97	4.18	2.62	1.20	0.88	0.98	0.65	0.96	1.62	0.83	0.57	1.37	1.58	0.87	0.67	1.48
STGCNN+Ours	0.91	1.10	1.24	1.19	1.79	0.97	0.89	0.61	1.10	0.36	0.65	0.51	0.95	0.43	0.77	0.69	1.15	0.31	0.74	0.66	0.85
T-GNN	1.74	1.01	1.45	1.04	2.19	1.28	2.68	1.84	1.34	1.11	0.99	0.69	1.25	1.74	0.97	0.53	1.18	1.70	0.88	0.79	1.31
T-GNN+Ours	0.91	0.96	1.29	1.10	1.27	0.94	1.26	0.76	1.16	0.40	0.79	0.61	1.16	0.37	0.74	0.61	1.06	0.42	0.74	0.70	0.89

Comparative Experiments

We compare our method to naive augmentations which not leverage any target information. Here are ADE/FDE results of 5 selected groups:

Method	A2B	B2C	C2D	D2E	E2A	AVG
STGCNN	1.12/1.49	1.18/1.97	0.54/0.98	0.45/0.57	0.93/1.37	0.84/1.28
+ Random	0.92/1.10	0.55/0.97	0.37/0.67	0.43/0.67	1.79/2.26	0.81/1.13
+ Ours	0.80/0.91	0.55/0.97	0.36/0.65	0.43/0.69	0.70/1.15	0.57/0.87

Visualizations

Method workflow

Statistics

Cases analysis

• Different colors represent different trajectories.
• Past trajectories are represented by dashed lines;
• Predict trajectories are shown as solid lines;
• ground truth future trajectories are depicted using gray dashed lines.

The source domain and the target domain have completely opposite road structures. Without our method, the model would still predict based on the direction of the source domain. Meanwhile, our method successful transfer of knowledge to the target domain.

Case3 and Case4 show group walking pattern and conflict walking pattern.