FlexiNet: An Adaptive Feature Synthesis Network for Real-Time Ego Vehicle Speed Estimation

Abstract

Ego vehicle speed estimation is critical for autonomous driving and advanced driver-assistance systems (ADAS), but traditional methods often fail in accuracy and computational efficiency under dynamic conditions. To address these challenges, we propose FlexiNet, a novel adaptive feature synthesis network that leverages monocular camera data to perform real-time speed estimation. FlexiNet integrates five key components, the Contextual Motion Analysis Block, Adaptive Feature Transformer, Spatial Feature Extraction Module, Motion Feature Extraction Module, and Dynamic Integration Gate, to effectively extract and fuse spatial and temporal features, thereby overcoming limitations of previous approaches by mitigating noise and capturing subtle motion dynamics. Comprehensive evaluations on the KITTI and nuImages datasets demonstrate FlexiNet’s superior performance. On the nuImages dataset, our model achieves an RMSE of 1.1358 m/s and an MAE of 0.9599 m/s, while on the KITTI dataset it records an RMSE of 1.9542 m/s and an MAE of 1.0610 m/s—reductions in error of up to 27.6% and 75.5% compared to baseline methods. These results validate the technical soundness and real-time capability of FlexiNet for deployment on embedded automotive platforms. By addressing critical gaps in previous research, FlexiNet makes a significant contribution toward the development of safer and more efficient autonomous vehicle technologies. The source code for FlexiNet is publicly available at here https://github.com/Geekgineer/FlexiNet