The ultrasonic echo signals from pavement coverings (water/ice/snow) typically exhibited low signal-to-noise ratio (SNR), making it challenging to accurately extract the onset point and thus affecting the time-of-flight (ToF) estimation precision. To address this issue, a total variation and Bayesian-based adaptive Kalman filtering (TV-BAKF) method was proposed. The method employed total variation analysis to assess noise distribution within a sliding window, deriving a differential factor that reflected both noise levels and signal characteristics. Combined with Bayesian estimation, it dynamically adjusted the process noise covariance and measurement noise covariance. Simulation tests were conducted and comparative analysis was performed with other filtering methods (improved wavelet transform, improved Savitzky-Golay (S-G) filter, Adaptive Unscented Kalman Filter (AUKF), and Variational Bayesian Adaptive Kalman Filter (VBAKF)), and the results showed that the TV-BAKF method maintained signal smoothness while achieving 0.915 rad phase fidelity and 0.917 amplitude fidelity, respectively, with favorable noise removal performance and a mean square error below 0.342. Testing on five types of coverings (water, ice, slush, wet snow, and dry snow) showed average depth detection errors below 0.55 mm for ice, water, and slush and no more than 0.96 mm for wet snow and dry snow, confirming its high-precision ToF estimation capability for low-SNR conditions .