🤖 machine learning

Learning-Based Speed Estimation from Accelerometer-Only Inertial Sensing

The paper introduces CarSpeedNet, a lightweight deep learning model that accurately estimates vehicle speed using only smartphone accelerometer data by leveraging a 4-second temporal window to achieve a root-mean-square error of 1.8 m/s without requiring gyroscope, wheel-odometry, or positioning inputs.

Original authors: Barak Or

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Barak Or

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Most of us take the speedometer for granted, a simple gauge that tells us how fast we are moving. In modern cars, this number comes from reliable sources like wheel sensors, the vehicle's internal computer network, or satellite navigation systems. But what if those systems are unavailable, or if the device trying to measure speed is just a smartphone sitting loosely in a cup holder? A smartphone's accelerometer, the sensor that detects motion, is notoriously difficult to use for this task on its own. It measures forces, not speed, and its readings are easily confused by the phone's orientation, the driver's hand movements, and the constant vibration of the road. Without a gyroscope to help stabilize the view or a connection to the car's own data, simply adding up these shaky acceleration numbers leads to a mess of errors almost instantly.

This is the challenge that researchers at ArtificialGate Ltd. set out to solve. They asked a fundamental question: can a machine learning model learn to guess a car's speed using only the raw, three-axis shaking data from a smartphone's accelerometer? They did not want to rely on the phone knowing which way is up, nor did they want to use any other sensors or external positioning data during the actual measurement. Instead, they proposed that the pattern of vibrations over a short period of time contains enough hidden clues to estimate speed. The core idea is that the way a car shakes is tied to how fast it is moving, and if a computer can learn to recognize those specific patterns, it can infer the speed without ever needing to calculate it directly from the forces.

The team, led by Barak Or, developed a system they call CarSpeedNet. To train this system, they collected over thirteen hours of driving data on real roads in Israel, using a Samsung Galaxy smartphone. They drove through both highways and city streets, recording the raw acceleration data at a high rate while simultaneously recording the true speed from a GPS system to serve as a reference. The researchers then fed this data into their model, teaching it to look at a window of past acceleration samples and predict the current speed. Crucially, they treated the length of this time window not just as a setting for the software, but as a fundamental part of the sensing problem itself. They reasoned that a longer window gives the computer more history to work with, allowing it to see more complete patterns of vibration, much like listening to a longer snippet of a song helps you recognize the tune.

The results showed that the length of this time window matters immensely. When the model looked at only a quarter of a second of data, its guesses were quite rough, with an average error of about 2.5 meters per second. However, as the researchers increased the window to four seconds, the accuracy improved dramatically. With that longer view, the average error dropped to just 0.72 meters per second. This improvement happened consistently as the window grew, suggesting that the extra time allowed the model to capture more of the rhythmic vibrations that are unique to specific speeds. The model itself is relatively compact, containing fewer than 180,000 adjustable settings, which means it could run efficiently on a standard smartphone without needing heavy computing power.

To ensure their approach was sound, the researchers compared CarSpeedNet against five other common types of time-based models, all using the same one-second window of data. CarSpeedNet performed as well as the best of these alternatives, matching the accuracy of a complex model inspired by image-recognition networks while using fewer computational resources. The study also highlighted a trade-off inherent in this method: while a longer window provides more information and leads to better accuracy, it also means the system is looking further back in time. This creates a slight delay, or "observation horizon," where the speed estimate is based on a history that is a few seconds old rather than the exact instant. For a driver, this is a small price to pay for a system that works without needing the car's internal sensors or a perfect GPS signal.

The study concludes that it is possible to estimate vehicle speed from a smartphone's accelerometer alone, provided the system is given enough time to observe the motion. The researchers did not claim this is a perfect solution for every situation; they noted that their tests were limited to the specific cars and phones used in their dataset, and they did not test how well the system would work if the phone were placed in a different spot or if the car were different. However, the findings offer a clear path forward. By treating the history of motion as a vital clue, rather than just noise to be filtered out, it is possible to turn a simple, low-cost sensor into a reliable speed estimator. This work suggests that even without the sophisticated hardware found in modern vehicles, the raw data from a phone, when viewed through the right lens, holds the key to understanding how fast we are moving.

Technical Summary: Learning-Based Speed Estimation from Accelerometer-Only Inertial Sensing

Problem Statement
Traditional vehicle speed estimation relies on wheel encoders, vehicle networks, or satellite navigation (GNSS), which are often inaccessible to low-cost auxiliary systems. While smartphone accelerometers are ubiquitous, direct integration of acceleration data to estimate speed is prone to rapid corruption by bias, mounting orientation errors, and vibration. Previous learning-based approaches have attempted to mitigate these issues by combining accelerometers with gyroscopes, utilizing stop-based recalibration, or employing complex alignment and attention mechanisms. However, a robust method for estimating scalar vehicle speed using only three-axis smartphone acceleration, without gyroscope, wheel-odometry, or positioning inputs at inference, remains a challenge due to the partial observability of velocity from short acceleration sequences.

Methodology
The paper introduces CarSpeedNet, a recurrent–convolutional regression model designed to infer vehicle speed from a finite window of three-axis acceleration samples.

Sensing Model: The system treats the problem as a finite-horizon mapping where the estimated speed $\hat{s}_t$ is a function of a sequence of acceleration samples $a_{t-W+1:t}$ within a window of length $W$ . The model does not attempt to separately identify orientation, bias, or road excitation but learns correlations directly from the training corpus.
Finite-Context Analysis: A core theoretical contribution is the characterization of the input window length as a sensing parameter rather than merely a network hyperparameter. The author derives that for nested histories, the minimum Bayes mean-square error is non-increasing with context length. Furthermore, they establish a "cue-coverage" relation: for a vibration component with spatial wavelength $\lambda$ and vehicle speed $s$ , the number of cycles $C_W$ presented to the network is proportional to the window length $W$ . This links the physical scale of road-induced vibrations to the temporal evidence available to the network.
Architecture: CarSpeedNet consists of a 100-unit bidirectional LSTM followed by four LSTM layers (50, 20, 20, 20 units) with batch normalization. It incorporates three 1D convolutional layers (64, 64, 32 filters) and a final 32-unit fully connected layer. The model contains 178,169 parameters (approx. 0.68 MiB for 32-bit weights).
Data: The study utilizes 13.2 hours of on-road driving data collected in Israel using a Samsung Galaxy smartphone. Data was acquired at 500 Hz, low-pass filtered, and downsampled to 20 Hz. The dataset is split into 10.3 hours for training, 2.6 hours for validation, and a 0.5-hour holdout set. GPS speed served as the reference for training and evaluation but was not used during inference.

Key Contributions

Accelerometer-Only Inference: The demonstration of a learning-based model capable of estimating scalar speed using only three-axis accelerometer data, excluding gyroscope or GNSS inputs at inference time.
Finite-Context Characterization: A theoretical framework treating window length as part of the sensing problem. The paper proves that longer nested histories cannot increase the minimum Bayes error and derives the relationship between window length, vehicle speed, and the number of vibration cycles available for estimation.
Comprehensive Benchmarking: A comparison of CarSpeedNet against five temporal-network alternatives (DNN, conventional LSTM, WaveNet, bidirectional LSTM, and a ResNet-inspired model) at a 1-second window, and an analysis of CarSpeedNet's performance across six different window lengths (0.25s to 4s).

Results
Performance was evaluated on the 0.5-hour holdout set using Root-Mean-Square Error (RMSE) and Mean Absolute Error (MAE).

Model Comparison (1-second window): At a 1-second input window, CarSpeedNet achieved an RMSE of 2.9 m/s and an MAE of 1.3 m/s. This performance was comparable to a ResNet-inspired alternative (2.9 m/s RMSE, 1.4 m/s MAE). The author notes that without repeated-seed results or a common tuning budget, statistical ranking between these top performers is not supported.
Temporal Context Analysis: Increasing the input window length yielded significant improvements. As the window grew from 0.25 s (5 samples) to 4 s (80 samples):
- RMSE decreased monotonically from 4.9 m/s to 1.8 m/s (a 63.3% reduction).
- MAE decreased monotonically from 2.50 m/s to 0.72 m/s (a 71.2% reduction).
- At the optimal 4-second window, the model achieved an RMSE of 1.8 m/s and an MAE of 0.72 m/s.
Observation Horizon: The author observes that while longer windows provide more repeated temporal evidence and reduce error, they inherently increase the observation horizon (latency) required for each prediction.

Significance and Claims
The paper modestly claims that the primary contribution is the finite-context characterization of the acceleration window. It posits that window length is not just a network setting but a fundamental part of the sensing problem, determining both the information set available to the estimator and the number of physical vibration cycles exposed to the network.

The empirical results confirm that longer nested histories improve estimation accuracy by providing a larger information set and more temporal evidence, consistent with the derived Bayes-optimal bounds. The study concludes that while CarSpeedNet offers a viable route for speed estimation without external sensors, the trade-off between accuracy and responsiveness (observation horizon) is a critical factor in system design. The author explicitly states that the evaluation is limited to the recorded corpus and does not include cross-device or cross-vehicle generalization studies, noting that robustness to different placements and environments remains a subject for future work.

Technical Summary: Learning-Based Speed Estimation from Accelerometer-Only Inertial Sensing

More like this