TinyOdom

Abstract

Deep inertial sequence learning has shown promising odometric resolution over model-based approaches for trajectory estimation in GPS-denied environments. However, existing neural inertial dead-reckoning frameworks are not suitable for real-time deployment on ultra-resource-constrained (URC) devices due to substantial memory, power, and compute bounds. Current deep inertial odometry techniques also suffer from gravity pollution, high-frequency inertial disturbances, varying sensor orientation, heading rate singularity, and failure in altitude estimation. In this paper, we introduce TinyOdom, a framework for training and deploying neural inertial models on URC hardware. TinyOdom exploits hardware and quantization-aware Bayesian neural architecture search (NAS) and a temporal convolutional network (TCN) backbone to train lightweight models targetted towards URC devices. In addition, we propose a magnetometer, physics, and velocity-centric sequence learning formulation robust to preceding inertial perturbations. We also expand 2D sequence learning to 3D using a model-free barometric g-h filter robust to inertial and environmental variations. We evaluate TinyOdom for a wide spectrum of inertial odometry applications and target hardware against competing methods. Specifically, we consider four applications: pedestrian, animal, aerial, and underwater vehicle dead-reckoning. Across different applications, TinyOdom reduces the size of neural inertial models by 31× to 134× with 2.5m to 12m error in 60 seconds, enabling the direct deployment of models on URC devices while still maintaining or exceeding the localization resolution over the state-of-the-art. The proposed barometric filter tracks altitude within ±0.1m and is robust to inertial disturbances and ambient dynamics. Finally, our ablation study shows that the introduced magnetometer, physics, and velocity-centric sequence learning formulation significantly improve localization performance even with notably lightweight models.

Systematic Review on Neural Architecture Search

“…It is easy to find BOs that are being used for different hardware and IoT devices for instance robotic and navigation applications [20]. They implemented BO and a temporal convolutional network (TCN) backbone to satisfy their goal of finding a model for an edge device with lower latency compared to man-made models.…”

“…Besides that, NAS can be adjusted to generate the most accurate, lightest, and/or fastest models for image classification [17], object detection [16], or semantic segmentation [18]. Furthermore, the NAS algorithms can optimize models for Internet of Things (IoT) devices as well as Microcontroller Unit (MCU)s [19,20]. However, NAS requires a computing power that is not accessible to all industries and people.…”

Deep Learning for Inertial Positioning: A Survey

“…This is illustrated in Figure 2 (b). By producing more accurate [34] Pedestrian, Trolley LSTM SL location displacement RIDI [35] Pedestrian SVM, SVR SL velocity for inertial data calibration Cortes et al [36] Pedestrian ConvNet SL velocity to constrain system drifts Wagstaff et al [37] Pedestrian LSTM SL zero-velocity detection for ZUPT Chen et al [38] Pedestrian, Trolley LSTM TL location displacement AbolDeepIO [39] UAV LSTM SL location displacement RINS-W [40] Vehicle RNN SL zero-velocity dection for KF Feigl et al [41] Pedestrian LSTM SL walking velocity Wang et al [42] Pedestrian LSTM SL walking heading for ZUPT Yu et al [43] Pedestrian ConvNet SL adaptive zero-velocity detection TLIO [44] Pedestrian ConvNet SL 3D displacement and uncertainty for EKF LIONet [45] Pedestrian Dilated ConvNet SL lightweight inertial model RoNIN [46] Pedestrian LSTM, TCN SL velocity for inertial data calibration Brossard et al [47] Vehicle ConvNet SL co-variance noise for KF StepNet [48] Pedestrian ConvNet, LSTM SL dynamic step length for PDR Wang et al [49] Pedestrian ConvNet SL measurement noise for Kalman Filter ARPDR [50] Pedestrian TCN SL stride length and walking heading for PDR IDOL [51] Pedestrian LSTM SL device orientation and location PDRNet [52] Pedestrian ConvNet SL step length and heading for PDR Buchanan et al [53] Legged Robot ConvNet SL integrate location displacement with leg odometry Zhang et al [54] Vehicle, UAV RNN SL independent motion terms Gong et al [55] Pedestrian LSTM SL fusing inertial data from two devices NILoc [56] Pedestrian ConvNet SL inertial relocalization RIO [57] Pedestrian DNN UL rotation-equivariance as supervision signal Wang et al [58] Pedestrian DNN SL efficient and low-latent model TinyOdom [59] Pedestrian, Vehicle TCN+NAS SL deployment on resource-constrained device CTIN [60] Pedestrian Transformer SL velocity and trajectory prediction DeepVIP [61] Vehicle ConvNet, LSTM SL velocity and heading for car localization Bo et al…”

“…When deploying deep learning-based inertial navigation on real-world devices, prediction accuracy and model efficiency must be considered. To address this, TinyOdom [59] aims to deploy neural inertial odometry models on resourceconstrained devices. It proposes a lightweight model based on temporal convolutional networks (TCN) [76] to learn position displacement and optimizes the model through neural architecture search (NAS) [77] to reduce model size between 31 and 134 times.…”

Dynamical Variational Autoencoders and Kalmannet: New Approaches to Robust High-Precision Navigation

“…Built upon state space models (SSM), this variant of the Kalman filter relies on conventional signal processing techniques, which heavily depend on manually designed simple mathematical models derived from domain expertise and assume Gaussian characteristics in the random models (Shlezinger, 2023). While the Kalman filter offers advantages like a compact footprint, minimal delay, and low power consumption, it encounters challenges when dealing with nonlinear state models and non-Gaussian random models (Yan et al, 2022, Das et al, 2015, Saha et al, 2022.…”