Neural Kalman Filter for Lateral Velocity Estimation

Built a Neural Kalman Filter with a TCN + Attention backbone to estimate vehicle lateral velocity (Vy) for the TI TDA4VM embedded platform, fusing the Ackermann formula and IMU integration via a learned Kalman gain K[t]. Pretrained the backbone on CarMaker simulation data, then fine-tuned on real-world data using LoRA to close the sim-to-real gap under limited real data. Reduced Vy estimation error vs. VSE (Ackermann-based baseline) by 48.6% RMSE (0.052 → 0.027 m/s) and 47.1% MAE (0.038 → 0.020 m/s), with the largest gains on low-μ surfaces and at high speeds where Ackermann-based estimation fails. Deployed the optimized model directly on TI TDA4VM via TIDLRT (EdgeAI TIDL tools).