International Conferences
| Dashcam-Based Ego-Vehicle Speed Estimation via Lane-Aware Spatiotemporal Learning |
| Year |
2026 |
| Month |
|
| Journal |
2026 / International Conference on Artificial Intelligence in Information and Communication (ICAIIC) |
| Author |
Woong-Chan Byun, Seung-Hyun Song, Chan-Bin Lim, Dong-Hee Paek, and Seung-Hyun Kong* |
|
Dashcams have become widely adopted and are utilized as evidentiary data for traffic accident analysis and collision reconstruction. In particular, ego-vehicle speed is a critical variable in accident investigations, as it is directly related to braking-distance analysis and speeding assessment. However, estimating ego-vehicle speed from dashcam video requires con- verting the motion observed in the image plane into the real- world domain. This conversion process relies on accurate intrinsic and extrinsic camera parameters. Such requirements are difficult to satisfy for commercial dashcams, which typically exhibit significant lens distortion, do not provide accessible calibration parameters, and are installed with user-specific positions and orientations. To address these limitations, this paper proposes a spatiotemporal learning framework that estimates ego-vehicle speed using only monocular dashcam video, without relying on any geometric camera information. The proposed framework integrates a ResNet-based feature extractor with a ConvLSTM module to model temporal motion patterns. In addition, lane segmentation is incorporated as an auxiliary task to provide geometric priors associated with lane structure and road scale. Experimental results on real-world driving datasets demonstrate that the proposed method reduces RMSE by 22.1% compared with state-of-the-art approaches, while requiring no camera calibration parameters. |
