Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences
1. Introduction
This project is for paper Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences. It concerns the single object tracking (SOT) of objects in point cloud sequences.
The input to the algorithm is the starting location (in the form of a 3D bounding box) of an object and the point cloud sequences for the scene. Our tracker then (1) provides the bounding box on each subsequent point cloud frame, (2) gets the dense shapes by aggregating the point clouds along with tracking.We also explore the usages on other applications, such as simulating LiDAR scans for data augmentation.
Please check our youtube video below for a 1-minute demonstration, and this link to the bilibili version.
This README
file describes the most basic usages of our code base. For more details, please refer to:
- Data Preprocessing: It describes how to convert the raw data in Waymo dataset into more handy forms, which can be used by our algorithms.
- Benchmark: It explains the selection of tracklets and construction of our benchmark. Note that the benchmark information is already in the
./benchmark/
and you may directly use it. The code in this part is for the purpose of verification. - Design: This documentation explains our design for the implementation. Reading this would be useful for understanding our tracker implementation and modifying it for your own purpose.
- Model Configs: We use the
config.yaml
to specify the behaviour of the tracker. Please refer to this documentation for detailed explanation. - Toolkit: Along this with project, we also provide several code snippets for visualizing the tracking results. This file discusses these toolkits we have created.
2. SOT API and Inference
2.1 Installation
Our code has been thoroughly tested using the environment of python=3.6
. For more detailed dependencies, please refer to the Environment
section below.
We wrap the usages of our code into a library sot_3d
, and the users may install it via the following command. The advantage of this installation command is that the behaviors of sot_3d
will keep synchronized with your modifications.
pip install -e ./
2.2 Tracking API
The main API tracker_api
is in main.py
. In the default case, it takes the model configuration, the beginning bounding box, and a data loader as input, output the tracking result as specified below. Some additional guidelines on this API are:
data_loader
is an iterator reading the data. On each iteration, it returns a dictionary, with the keyspc
(point cloud) andego
(the transformation matrix to the world coordinate) as compulsory. An example ofdata_loader
is in example_loader.- When you want to compare the tracking results with the ground truth along with tracking, please provide the input argument
gts
and import the functioncompare_to_gt
, the data typesot_3d.data_protos.BBox
. The gts are a list ofBBox
. - We also provide a handy tool for visualization. Please import
from sot_3d.visualization import Visualizer2D
andframe_result_visualization
for a frame-level BEV visualization.
import sot_3d
from sot_3d.data_protos import BBox
from sot_3d.visualization import Visualizer2D
def tracker_api(configs, id, start_bbox, start_frame, data_loader, track_len, gts=None, visualize=False):
"""
Args:
configs: model configuration read from config.yaml
id (str): each tracklet has an id
start_bbox ([x, y, z, yaw, l, w, h]): the beginning location of this id
data_loader (an iterator): iterator returning data of each incoming frame
track_len: number of frames in the tracklet
Return:
{
frame_number0: {'bbox0': previous frame result, 'bbox1': current frame result, 'motion': estimated motion}
frame_number1: ...
...
frame_numberN: ...
}
"""
2.3 Evaluation API
The API for evaluation is in evaluation/evaluation.py
. tracklet_acc
and tracklet_rob
compute the accuracy and robustness given the ious in a tracklet, and metrics_from_bboxes
deals with the cases when the inputs are raw bounding boxes. Note that the bounding boxes are in the format of sot_3d.data_protos.BBox
.
def tracklet_acc(ious):
...
""" the accuracy for a tracklet
"""
def tracklet_rob(ious, thresholds):
...
""" compute the robustness of a tracklet
"""
def metrics_from_bboxes(pred_bboxes, gts):
...
""" Compute the accuracy and robustness of a tracklet
Args:
pred_bboxes (list of BBox)
gts (list of BBox)
Return:
accuracy, robustness, length of tracklet
"""
3 Building Up the Benchmark
Our LiDAR-SOT
benchmark selects 1172 tracklets from the validation set of Waymo Open Dataset. These tracklets satisfy the requirements of mobility, length, and meaningful initialization.
The information of selected tracklets is in the ./benchmark/
. Each json
file stores the ids, segment names, and the frame intervals for each selected tracklet. For replicating the construction of this benchmark, please refer to this documentation.
4. Steps for Inference/Evaluation on the Benchmark
4.1 Data Preparation
Please follow the guidelines in Data Preprocessing. Suppose your root directory is DATA_ROOT
.
4.2 Running on the benchmark
The command for running on the inference is as follows. Note that there are also some other arguments, please refer to the main.py
for more details.
python main.py \
--name NAME \ # The NAME for your experiment.
--bench_list your_tracklet_list \ # The path for your benchmark tracklets. By default at ./benchmark/bench_list.json.
--data_folder DATA_ROOT \ # The location to store your datasets.
--result_folder result_folder \ # Where you store the results of each tracklet.
--process process_number \ # Use mutiple processes to split the dataset and accelerate inference.
After this, you may access the result for tracklet ID
as demonstrated below. Inside the json files, bbox0
and bbox1
indicates the estimated bounding boxes in frame frame_index - 1
and frame_index
.
-- result_folder
-- NAME
-- summary
-- ID.json
{
frame_index0: {'bbox0': ..., 'bbox1': ..., 'motion': ...,
'gt_bbox0': ..., 'gt_bbox1': ..., 'gt_motion': ...,
'iou2d': ..., 'iou3d': ...}
frame_index1: ...
frame_indexN: ...
}
4.3 Evaluation
For computing the accuracy and robustness of tracklets, use the following code:
cd evaluation
python evaluation.py \
--name NAME \ # the name of the experiment
--result_folder result_folder \ # result folder
--data_folder DATA_ROOT \ # root directory storing the dataset
--bench_list_folder benchmark_list_folder \ # directory for benchmark tracklet information, by default the ./benchmark/
--iou # use this if already computes the iou during inference
--process process_number # use multiprocessing to accelerate the evaluation, especially in cases of computing iou
For the evaluation of shapes, use the following code:
cd evaluation
python evaluation.py \
--name NAME \ # the name of the experiment
--result_folder result_folder \ # result folder
--data_folder DATA_ROOT \ # root directory storing the dataset
--bench_list_folder benchmark_list_folder \ # directory for benchmark tracklet information, by default the ./benchmark/
--process process_number # Use mutiple processes to split the dataset and accelerate evaluation.
5. Environment
This repository has been tested and run using python=3.6
.
For inference on the dataset using our tracker, the following libraries are compulsory:
numpy, scikit-learn, numba, scipy
If the evaluation with ground-truth is involved, please install the shapely
library for the computation of iou.
shapely (for iou computation)
The data preprocessing on Waymo needs.
waymo_open_dataset
Our visualization toolkit needs.
matplotlib, open3d, pangolin
6. Citation
If you find our paper or repository useful, please consider citing
@article{pang2021model,
title={Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences},
author={Pang, Ziqi and Li, Zhichao and Wang, Naiyan},
journal={arXiv preprint arXiv:2103.06028},
year={2021}
}