PeDistNet: AI-based Indoor Navigation for Pedestrians
Revolutionizing Distance Estimation for Pedestrian using Smartphone: A Deep Dive into Advanced LSTM and Convolutional Model Architecture with SOTA performance. Walking while smartphone in the pocket for 493 meters with only 0.06% error.
Introduction
In the era of artificial intelligence and data science, the integration of AI with existing smartphone sensors is revolutionizing indoor navigation for pedestrians. This transformation is illustrated through the utilization of “OxIOD: The Dataset for Deep Inertial Odometry.” This dataset is used for the evolution of our suggested AI-based pedestrian navigation systems, allowing users to know accurately their traveled distance, also in indoor environments – without satellite availability (no GPS).
Watch it on MetaOr YouTube Channel:
The Role of OxIOD Dataset in Pioneering Indoor Navigation
OxIOD, representing the Oxford Inertial Odometry Dataset, is a curated dataset, emblematic of real-world scenarios in indoor navigation. This dataset comprises data recorded from an iPhone 7P, carried by a pedestrian in various common situations like handheld, in the pocket, in a handbag, or on a trolley. The pocket data capturing 11 sequences over 5,622 seconds (approximately 1.5 hours) and covering a total distance of 4.2 km, OxIOD provides an authentic reflection of daily smartphone usage.
Detailed Dataset Characteristics
- Device Used: iPhone 7P.
- Data Division: The dataset is divided into a training set covering 3,833 meters and a testing set spanning 493 meters. In minutes: 83 minutes for training and 10 minutes for test.
- Downsampling Approach: The original data, recorded at 100Hz, was downsampled to 20Hz. This was important for reducing noise and ensuring the efficient processing of the model.
Our Suggested AI Model: PeDistNet
Our methodology involved developing the PeDistNet (pedestrian distance network)deep learning-based model. This approach shifts from the traditional step detection and step length estimation approaches to a more sophisticated analysis of the relationship between IMU signals and the distance traveled. The motivation here is to sum up small distances in short time windows, similar to the way of calculating the length of an arc (you may read also my previous work LenghtNet Paper, Post):
Dataset and general setting
We defined our task as a supervised learning regression problem, aiming to predict the distance traveled in 10-second intervals using IMU sensor data. The model processes inputs of 6x200 inertial sensor data points over 10 seconds, with the output being the scalar value of the traveled distance in that duration. The inertial sensor data refers to the accelerometer and gyroscopes MEMS sensors, existing in any modern smartphone device, and measures accelerations and angular velocity of the smartphone device (and the pedestrian who holds it).
The ground truth distance was calculated using GPS records. It was recorded outdoors to achieve the ground truth of the distance, using a smartphone device, equipped with a GPS receiver, accelerometer, gyroscope, and a recording app for this data. It is important to synchronize between all sensors so all readings will be considered at the same timeline without any delay ( OxIOD Dataset contains Sync data).
Architecture
Various Deep Neural Network (DNN) architectures were employed until we found the best model. The architecture is a sequential model, including four 1D convolutional layers for feature extraction, followed by a combination of bidirectional and standard LSTM layers with dropout and batch normalization for capturing temporal dependencies. The model concludes with a GlobalAveragePooling1D layer for dimensionality reduction and a dense linear output layer for continuous value prediction. This setup was trained for our task of distance estimation using sensor data.
Optimizer and Loss Function
The training process utilized the ADAM optimizer. The Mean Squared Error function was used as the loss function. The inputs are the accelerometer readings (a) and the gyroscope readings (w) in all three axes. Each input to the PeDistNet is of a window length corresponding to 10 seconds of data (20 Hz frequency — results in a window length of w=200).
The estimated distance was compared with the truth one (GPS) using the RMSE criterion:
Main Results
We explored many architectures until we received the best-performing model, which we called PeDistNet.
To evaluate the performance on the test set, all inputs were inserted into the trained model (10 seconds for each) and we calculated the error — as can be seen in the figure below:
The average RMSE for all examples in the test set is 12 [cm].
By summing all the PeDistNet outputs, we calculated the total traveled distance and compared it with the ground truth (GPS data) as shown in the figure below.
Resulting in only 34 centimeters of final distance error after 493.4 meters of walking (reflecting 0.06% on the test set).
Implications and Future Applications
The integration of AI with smartphone sensor data, as demonstrated by our application of the OxIOD dataset, marks a transformative step in indoor pedestrian navigation. Our model, PeDistNet, offers a novel and effective approach to distance estimation, extending beyond traditional methodologies. This advancement opens avenues for applications in health monitoring, such as tracking the total distance walked over a given period, and provides a benchmark for validating results in the indoor navigation domain.
Conclusion
The use of “OxIOD: The Dataset for Deep Inertial Odometry” for setting AI-based models like PeDistNet is redefining the field of indoor pedestrian navigation. By accurately interpreting the complex relationship between IMU signals and traveled distances, this capability sets a new standard in the precision and reliability of indoor navigation systems with widely available smartphone devices. The future of this field is poised for further breakthroughs, with extensive applications ranging from personal fitness to sophisticated navigation solutions in complex indoor environments.
In our future research, we will explore various input window sizes, test our results on a big database, and modify the model respectively. This initial phase results gave us the motivation to keep solving the entire pedestrian dead reckoning (PDR) problem with AI approaches.
About the Author
Dr. Barak Or is a professional in the field of artificial intelligence and sensor fusion. He is a researcher, lecturer, and entrepreneur who has published numerous patents and articles in professional journals. Dr. Or leads the MetaOr Artificial Intelligence firm. He founded ALMA Tech. LTD holds patents in the field of AI and navigation. He has worked with Qualcomm as DSP and machine learning algorithms expert. He completed his Ph.D. in machine learning for sensor fusion at the University of Haifa, Israel. He holds M.Sc. (2018) and B.Sc. (2016) degrees in Aerospace Engineering and B.A. in Economics and Management (2016, Cum Laude) from the Technion, Israel Institute of Technology. He has received several prizes and research grants from the Israel Innovation Authority, the Israeli Ministry of Defence, and the Israeli Ministry of Economic and Industrial. In 2021, he was nominated by the Technion for “graduate achievements” in the field of High-tech.
Website www.metaor.ai Linkedin www.linkedin.com/in/barakor/ YouTube www.youtube.com/channel/UCYDidZ8GUzUy_tYtxvVjRiQ
— -
References
[1] Chen, Changhao, et al. “Oxiod: The dataset for deep inertial odometry.” arXiv preprint arXiv:1809.07491 (2018).
[2] Chen, Changhao, et al. “Deep-learning-based pedestrian inertial navigation: Methods, data set, and on-device inference.” IEEE Internet of Things Journal 7.5 (2020): 4431–4441.
