This study proposes a parallelization-based asynchronous dual-camera frame acquisition structure to improve the temporal resolution of camera-based perception data for autonomous mobility control. In a conventional single-camera perception system, the...
This study proposes a parallelization-based asynchronous dual-camera frame acquisition structure to improve the temporal resolution of camera-based perception data for autonomous mobility control. In a conventional single-camera perception system, the data update period is limited by the physical frame rate of the camera, which may cause delays in lane recognition, road-curvature estimation, and control input updates. To address this limitation, two 30 FPS USB cameras were operated asynchronously with a time offset of approximately 0.5 frame instead of capturing images simultaneously. The frames acquired from the two cameras were subsequently arranged and integrated according to their timestamps.
To verify the effectiveness of the proposed structure, basic experiments were conducted using desktop PC and mini PC environments. The frame interval, FPS, and communication-period stability of the two cameras were compared. The experimental results showed that both platforms achieved an average communication time difference of approximately 17.15 ms, which was nearly half the frame interval of a single 30 FPS camera, approximately 33.3 ms. The desktop PC exhibited a narrower frame-interval distribution and more stable FPS characteristics than the mini PC, indicating that the temporal stability of the asynchronous dual-camera structure may be affected by the processing performance and scheduling characteristics of the computing platform.
The proposed structure was subsequently applied to a 1/10 scale mobility platform and evaluated in a real driving environment. In the driving experiment, cam0 and cam2 independently acquired lane images, and the measurements obtained from both cameras were arranged in timestamp order to generate the integrated cam0+cam2 data. After bird’s-eye-view transformation and vehicle-center-based coordinate correction, the detected left and right lane points were used to determine the lane center points. The lane centerline was then approximated using a third-order polynomial, and the polynomial coefficients a, b, c, and d were calculated. Road curvature, denoted by , was calculated from the first and second derivatives of the estimated polynomial at a predefined evaluation position.
To reduce instantaneous variations in the polynomial coefficients caused by vehicle vibration, uneven road surfaces, camera motion, and changes in the detected lane points, a random walk Kalman filter was applied. The measurements from cam0 and cam2 were not simply added or averaged; instead, they were sequentially entered into a single Kalman filter according to their timestamps. The filtered polynomial coefficients were subsequently used to calculate the curvature of the cam0+cam2 result.
The driving data were divided into straight and curved sections, and the FPS, polynomial coefficients, and curvature values of cam0, cam2, and cam0+cam2 were compared. In both sections, the cam0+cam2 data provided a higher effective update rate than either individual camera. In the straight section, the curvature values remained within a relatively consistent range, although temporary fluctuations occurred due to vehicle motion and lane-detection changes. The integrated cam0+cam2 result exhibited a narrower distribution and smoother temporal variation than the individual camera results, indicating that instantaneous measurement fluctuations were partially reduced.
In the curved section, the curvature values changed more distinctly than in the straight section because the estimated lane centerline reflected the curved driving path. Cam0 and cam2 exhibited different curvature magnitudes owing to differences in camera mounting position, field of view, and residual coordinate-transformation errors. The curvature estimated from the cam0+cam2 data was generally distributed between the values obtained from the two individual cameras and showed comparatively smaller temporal variation. These results demonstrate that the proposed sequential integration method can provide more temporally continuous polynomial coefficients and curvature information while preserving the geometric characteristics of the driving path.
Consequently, the proposed parallelization-based asynchronous dual-camera structure shortened the perception-data update period by interleaving actual frames acquired from two low-cost cameras along the time axis, without directly increasing the physical FPS of either camera or generating artificial interpolation frames. Its applicability was further demonstrated through third-order polynomial-based lane modeling and curvature estimation in an actual driving environment. Therefore, the proposed structure can serve as a practical and low-cost sensor configuration for improving the temporal resolution of perception data and supporting more frequent lane-state and control-input updates in autonomous mobility systems.