NeurIT: Pushing the Limit of Neural Inertial Tracking for Indoor Robotic Things

This repository contains the system set up for NeurIT, the self-built dataset, and the source code for TF-BRT.

1. System Setup

To collect robotic tracking data, my collaborator helped built up a customized robotic system. As shown in the following picture, the system is composed of six parts:

1) CH110 IMU

Collect raw IMU data, including accelerations, gyroscopes, and magnetometers. In addition, it provides the pose information of the robot.

2) Intel NUC

An Intel NUC with an i7-1260P CPU, 32GB memory, is used for running the whole system.

3) MID360 LiDAR

The hardware for FAST-LIO algorithm. FOV: 360°*59°; Point Cloud Density: 40-line.

4) LIVOX AVIA

A narrow-FOV LiDAR, which is used to demonstrate the fact that LiDAR-inertial tracking systems cannot work well in simple and plain indoor environments.

5) Tango Phone

A commercialized visual-inertial tracking system for comparison.

6) Tracer Mini

We use AgileX Tracer-Mini for the robot’s territory.

We build up a workspace for ROS to collect the IMU data and the ground truth (FAST-LIO). You can download the worksapce via this link System Setup. For instance, we costomize a ROS topic imu_all to publish the data we want to collect. You can follow the following file in the workspace for data collection: /mid360_ws/src/imu/serial_imu/scripts/subscriber.py

2. Dataset

To verify the effectiveness of our proposed TF-BRT, we collect NeurIT dataset. The dataset is collected across three buildings in the campus. The dataset is splited into four parts: training set, validation set, test seen set, and test unseen set.

Training, validation, and test seen sets are collected in Building A (floor 1 & floor 5). Test unseen set is collected in Building B and Building C. NeurIT dataset contains 110 sequences, totaling around 15 hours of tracking data that corresponds to a travel distance of about 33.7 km. Each sequence of data lasts 6∼10 minutes. You can download the dataset via this link NeurIT Dataset.

The file includes original_data, uniform_data, and data_gen_neurit.py. You can use data_gen_neurit.py file to convert original_data to the uniform_data. The data pre-processing methods include time synchronization, and coordinate system transformation. The original_data contains the raw data collected from IMU and the ground truth. The data format is given below:

synced
├── acce (acceleration in body frame)
├── gyro (gyroscope in body frame)
├── magnet (magnetometer measured in body frame)
├── rv (rotation vector in quaternion format)
├── time (time stamp: 200Hz)

pose
├── time (time stamp)
├── fastlio_pos (position of the robot in navigation frame)
├── fastlio_ori (orientation of the robot in navigation frame)

When you want to train TF-BRT, please use the data in uniform_data.

3. TF-BRT (Time-Frequency Block-Recurrent Transformer)

You could download or git clone this repository to test the performance of TF-BRT. It contains three steps for training and evaluation.

1. Configuration
2. Training
3. Testing

3.1 Configuration

All of the hyperparameters, model path, data path are defined in tf-brt/source/config.py for convenience, as you donnot have to type in anything in the command line.

Please download the NeurIT dataset to the folder, and fill in necessary path information before training. The following setting code is for your reference.

DATASET = "neurit" # Dataset name
MODEL_TYPE = "tf-brt"
DATA_DIR = './NeurIT Dataset/uniform_data/train_dataset' # Dataset directory for training
VAL_DATA_DIR = './NeurIT Dataset/uniform_data/val_dataset' # Dataset directory for validation
TEST_DIR = './NeurIT Dataset/uniform_data/test_seen' # Dataset directory for testing (test_seen & test_unseen)
OUT_DIR = './prediction_model/neurit/test1' # Output directory for both traning and testing

Please go through config.py file scrupulously before entering the training stage. Feel free to adjust the hyperparameters if you like.

3.2 Training

After everthing is done, you can type in python tf-brt.py --mode train in your command line for training. The checkpoints will be stored in the OUT_DIR defined in the config.py file.

3.3 Testing

In NeurIT dataset, we have test_seen and test_unseen datasets. For begin with, you need to change the setting for TEST_DIR, OUT_DIR, and MODEL_PATH.

The following setting code is for your reference (test_seen):

TEST_DIR = './NeurIT Dataset/uniform_data/test_seen' # Dataset directory for testing (test_seen & test_unseen)
OUT_DIR = './NeurIT/tf-brt/result/neurit/test1/seen' # Output directory for both traning and testing
MODEL_PATH = './NeurIT/tf-brt/prediction_model/neurit/test1/checkpoints/checkpoint.pt' # Model path for testing

Then, the testing results, including trajectory visualizations, and four evaluation metrics (ATE, RTE, PDE, AYE), will be recorded in the output file. Please refer to the paper for more details about the evaluation metrics.