Driving Behavior Classification Using a ConvLSTM

A hybrid ConvLSTM architecture that classifies normal and aggressive driving behavior from smartphone accelerometer, gyroscope and GPS data, benchmarked against Stacked LSTM and BiLSTM baselines.

Reading driving behavior from a phone in a pocket

Smartphones already carry the sensors needed to observe how a car is being driven, but the raw signal is noisy, irregularly sampled and has no notion of where one maneuver ends and the next begins.

• Classify driving behavior as normal or aggressive using only commodity smartphone sensors.
• Accuracy has to hold outside a controlled lab setting.
• No dedicated telematics hardware allowed in the loop.

Accelerometer, gyroscope and GPS streams

Input is limited to three smartphone sensor streams captured during real trips, each with its own scale, sampling rate and noise profile.

• Accelerometer — linear motion and braking intensity.
• Gyroscope — rotation, cornering and lane-change dynamics.
• GPS — trajectory and speed context.
• Labels: normal vs. aggressive driving behavior.

Preprocessing carried as much weight as the model

Before any training, the raw streams pass through a dedicated preprocessing stage that makes the signal comparable and gives the model coherent units to learn from.

• Normalization across sensors and trips.
• Noise filtering to remove sensor jitter.
• Maneuver segmentation into fixed windows, since boundaries are not labelled in the raw data.

Convolution for shape, LSTM for time

The ConvLSTM stacks two complementary ideas: a convolutional stage that extracts local sensor patterns inside each maneuver window, and a recurrent stage that models how those patterns evolve across the sequence.

• Convolutional layers learn spatial structure across the sensor axes.
• LSTM layers carry temporal dependencies between consecutive windows.
• A dense classification head outputs the normal / aggressive decision.

What the experiment settled

The hybrid architecture outperformed both sequence-only baselines trained on identical data, at the cost of a heavier training loop.

• Hybrid architectures win when the input has both spatial and temporal structure.
• Preprocessing quality mattered as much as the model choice.
• Benchmarking against simpler baselines is what makes an accuracy claim defensible in review.