Labeled 81 GHz FMCW radar data for classification between car and motorcycle

This dataset was created as part of a diploma thesis on radar-based object classification using machine learning techniques titled "Radar signal classification using machine learning techniques". It consists of annotated radar measurements collected under controlled conditions using multiple Texas Instruments FMCW radar modules (AWR1443 and AWR1843). The objective of the dataset is to support the development of reliable classification systems for automotive targets (cars vs motorcycles) based solely on radar data, without reliance on visual information (e.g., cameras or LiDAR). Data Collection Radar data were recorded from moving vehicles (4 cars and 4 motorcycles) coming towards a fixed radar platform at consistent speeds and trajectories. The sensors were positioned on a custom-built frame. Each recorded frame contains multiple radar point detections with the following attributes: Spatial coordinates (X, Y, Z) in meters Doppler (radial velocity) in m/s SNR (signal-to-noise ratio) A camera was also used during data collection, not for classification, but to enable manual annotation (labeling) of the radar data with ground truth labels (car or motorcycle). Preprocessing & Format Raw radar binary data were parsed, filtered, and synchronized into structured CSV files. The radar points from all the radars were merged into one CSV file. Clustering was performed using the DBSCAN algorithm, followed by Kalman filtering for multi-object tracking and unique ID assignment. Each tracked object (cluster) is stored per buffer (buffer of 5 frames) with corresponding features. Extracted features include geometric (e.g., eigenvalues, bounding box volume), motion (e.g., ΔX, ΔY, speed), and signal quality (e.g., Doppler/SNR statistics and entropy). Labeling was applied manually for supervised learning models. Use Cases The dataset is designed to train and evaluate ML models for radar-based target classification. It supports: Binary classification (car vs motorcycle) Exploratory data analysis (EDA) Dimensionality reduction (PCA, UMAP, t-SNE) Real-time inference systems for ADAS and Smart City applications File Contents Eight zip files, each one is a vehicle Clustered and labeled target data Extracted features Scripts for preprocessing, feature extraction, and model training [1] (available via GitHub upon request) [1] “Master/Ioannis at main · uniwa-radar/Master.” Accessed: Jul. 16, 2025. [Online]. Available: https://github.com/uniwa-radar/Master/tree/main/Ioannis