Academia.eduAcademia.edu

Figure 1 – uploaded by Sachini Herath

See full PDFdownloadDownload figure
Figure 1.1: Inertial navigation is the problem of estimating the position and orientation of a moving subject only from a sequence of IMU sensor data. We presents a new benchmark, new algorithms, and extensive evaluations of existing techniques for inertial navigation. with everyday smartphone usage scenarios. For example, an IMU must be attached to a foot to Most existing inertial navigation algorithms require unrealistic constraints that are incompatible

Figure 1 1: Inertial navigation is the problem of estimating the position and orientation of a moving subject only from a sequence of IMU sensor data. We presents a new benchmark, new algorithms, and extensive evaluations of existing techniques for inertial navigation. with everyday smartphone usage scenarios. For example, an IMU must be attached to a foot to Most existing inertial navigation algorithms require unrealistic constraints that are incompatible

Related Figures (24)

Figure 2.2: Kalman filter uses a recursive algorithm that predict the state and its uncertainty P at each time step, and correct the estimate and measurement noise models once the measurements are available 2
Figure 2.2: Kalman filter uses a recursive algorithm that predict the state and its uncertainty P at each time step, and correct the estimate and measurement noise models once the measurements are available 2
Figure 2.3: Euler angle representation: The axes of the original frame are denoted as x,y,z and the axes of the rotated frame as X,Y,Z. > denote rotation from coordinate frame S to target coordinate frame T as R?”.
Figure 2.3: Euler angle representation: The axes of the original frame are denoted as x,y,z and the axes of the rotated frame as X,Y,Z. > denote rotation from coordinate frame S to target coordinate frame T as R?”.
Figure 2.4: Any rotation can be represented as a rotating around axis é by angle @. This notion is used by Euler-Rodrigous vector and quaternion representations. *
Figure 2.4: Any rotation can be represented as a rotating around axis é by angle @. This notion is used by Euler-Rodrigous vector and quaternion representations. *
in certain cases. More precisely, stabilized IMU coordinate frame, shown in Fig. 2.5, is obtained
in certain cases. More precisely, stabilized IMU coordinate frame, shown in Fig. 2.5, is obtained
IONet dataset, namely OXIOD used a high precision motion capture system (Vicon) under four
IONet dataset, namely OXIOD used a high precision motion capture system (Vicon) under four
Figure 3.1: RoNIN data collection pipeline 3.1 Data Collection Pipeline
Figure 3.1: RoNIN data collection pipeline 3.1 Data Collection Pipeline
3.1.2. Two Device Protocol
3.1.2. Two Device Protocol
Table 3.2: Standard deviations of the yaw/roll/pitch angles and the speed [m/s], averaged over all the sessions in each dataset. 3.3. Data Description angle show the direction of motion. RoNIN dataset has 3 to 10 times more variation.
Table 3.2: Standard deviations of the yaw/roll/pitch angles and the speed [m/s], averaged over all the sessions in each dataset. 3.3. Data Description angle show the direction of motion. RoNIN dataset has 3 to 10 times more variation.
Figure 4.1: RoNIN Architectures: Neural network modules are in green and transformation layers are in white. training. IMU sensor measurements come from moving device coordinate frames (Fig. 4.2a.), whil ‘eature representations, in our case the choice of coordinate frames, have significant impacts on the
Figure 4.1: RoNIN Architectures: Neural network modules are in green and transformation layers are in white. training. IMU sensor measurements come from moving device coordinate frames (Fig. 4.2a.), whil ‘eature representations, in our case the choice of coordinate frames, have significant impacts on the
eround-truth motion trajectories come from a global coordinate frame. RoNIN uses a heading-
eround-truth motion trajectories come from a global coordinate frame. RoNIN uses a heading-
fable 5.1: Position evaluation: We compare five competing methods: Naive Double Integration(NDI), Pedestrian Dead Reckoning(PDR), RIDI, ONet, and RoNIN (3 variants) on three datasets: the RIDI dataset, the OXIOD dataset and our new dataset. The top three results are highlighted in ed, green, and blue colors per row.
fable 5.1: Position evaluation: We compare five competing methods: Naive Double Integration(NDI), Pedestrian Dead Reckoning(PDR), RIDI, ONet, and RoNIN (3 variants) on three datasets: the RIDI dataset, the OXIOD dataset and our new dataset. The top three results are highlighted in ed, green, and blue colors per row.
Figure 5.1: Selected visualizations. We select 3 examples from each dataset and visualize recon- structed trajectories from all competing methods. We choose RoNIN ResNet as our method, since its performance is more consistent across datasets. For each sequence we mark the trajectory length and report both ATE and RTE of our method. We also mark physical dimensions in all sequences to demonstrate that our method estimate trajectories with accurate scales. Three examples from our RoNIN dataset (left column) contains complex natural motions, where all other methods fail. Se- quences from the RIDI dataset (middle column) contains hard motions. In particular, the middle example of the middle column contains extensive backward motion, where our method handles el- egantly. Sequences from the OXIOD dataset (right column) are mostly short sequences with easy motions. However, our method gives large error for a few sequences (e.g. the bottom one) due to the large error in the provided device orientations.
Figure 5.1: Selected visualizations. We select 3 examples from each dataset and visualize recon- structed trajectories from all competing methods. We choose RoNIN ResNet as our method, since its performance is more consistent across datasets. For each sequence we mark the trajectory length and report both ATE and RTE of our method. We also mark physical dimensions in all sequences to demonstrate that our method estimate trajectories with accurate scales. Three examples from our RoNIN dataset (left column) contains complex natural motions, where all other methods fail. Se- quences from the RIDI dataset (middle column) contains hard motions. In particular, the middle example of the middle column contains extensive backward motion, where our method handles el- egantly. Sequences from the OXIOD dataset (right column) are mostly short sequences with easy motions. However, our method gives large error for a few sequences (e.g. the bottom one) due to the large error in the provided device orientations.
Figure 5.2: The ratio of RoNIN testing sequences under different thresholds on the two metrics. For instance, the left graph shows the ratio of sequences where the ATE is below a certain threshold.
Figure 5.2: The ratio of RoNIN testing sequences under different thresholds on the two metrics. For instance, the left graph shows the ratio of sequences where the ATE is below a certain threshold.
Figure 5.3: RTE metric with different window sizes for testing sequences of RoNIN dataset.
Figure 5.3: RTE metric with different window sizes for testing sequences of RoNIN dataset.
5.2.2 Body heading evaluation
5.2.2 Body heading evaluation
Figure 5.4: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Table 5.2: Body heading estimation task. We compare our RoNIN heading network with the base- line, which simply reports the device heading as the body heading.
Figure 5.4: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Table 5.2: Body heading estimation task. We compare our RoNIN heading network with the base- line, which simply reports the device heading as the body heading.
Figure 5.5: Selected visualization of heading angle estimations. Note that all three dimensions are needed and we rotate predicted velocities to the heading agnostic
Figure 5.5: Selected visualization of heading angle estimations. Note that all three dimensions are needed and we rotate predicted velocities to the heading agnostic
Figure A.1: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.1: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.2: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.2: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.3: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.3: Selected visualizations of trajectories from 3 variants of RONIN method on our dataset. Positional errors are marked within each figure, where "A" denotes ATE and "R" denotes RTE.
Figure A.4: Additional visualization of heading estimation.
Figure A.4: Additional visualization of heading estimation.
Related topics:
Computer ScienceInertial Measurement Unit
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved