A Lightweight Multi-Target Vehicle Detection and Tracking Method with Roadside Millimeter Wave Radar

Millimeter wave (MMW) radar is widely used for vehicle detection due to its high-precision speed measurements, low cost, the ability of providing depth information, and resist weather interference, e.g., fog, rain, and snow. Particularly, in the environment with extremely low illumination and visability, it can still operate normally. However, the performance of MMW radar-based sensing technology is limited by its sparse spatial resolution, which makes the recognition of the target point from the noise point even harder. Especially for multi-target detection and tracking, adjacent targets are often incorrectly recognized as one object. Even the deep learning-based data processing algorithms are proposed to solve the aforementioned issue, the excessive computational complexity and the requirements for large amount of training data also limit the value for implementation. Thus, this study presents a lightweight method to identify and track multiple vehicles simultaneously. The proposed method is a multi-step framework, including adapting the DBSCAN algorithm, tracking the vehicle with Hungarian algorithm and Kalman filter, and identifying vehicle types by radar cross-sections and L-shaped bounding boxes fitted by clustered point cloud. Adjustments to DBSCAN algorithm include creating automatically adaptive parameters based on point cloud distance and velocity, using elliptical range for nearest neighbor searches, incorporating speed into clustering features, and fusing multi-frame point clouds. According to the results of the demonstration based on 22,000 frames of MMW radar data, the modified DBSCAN algorithm eliminates 93% of the original data, which is estimated to be noise points. It is observed that the proposed method has significantly improved trajectory continuity, object recognition at extreme distances and concurrent adjacent trajectories recognition. Vehicle classification method by incorporating more effective features can be one of the future research directions.