The tunable diode laser absorption spectroscopy (TDLAS) has the narrow linewidth characteristic of tunable diode lasers, so the gas concentration measurement of high precision and selectivity can be achieved by selecting the single absorption line of the specific gas to eliminate the interference of other gases, which has wide applications in the gas concentration detection. However, under different application conditions and environments, the people need to solve corresponding technical problems in hardware and data processing. In this paper, the TDLAS spectral data processing problem in the telemetry system for the vehicle exhaust CO concentration has been mainly studied. The system telemetered the exhaust CO concentration of the driving vehicle with the echo signals from the road diffuse reflection. Because the echo signal of laser scanning spectral is affected by factors such as the variation of the diffuse reflection surface, the change of the air environment, and the influence of exhaust turbulence, the signal collected by the detector is not only weak but also mixed with various noises, which means the SNR of the measured optical path is comparatively weak, so an adaptive hierarchical Savitzky-Golay (S-G) smoothing filter algorithm has been proposed in this paper, which can realize the spectral filtering processing to inversion the CO concentration more accurately. The S-G filtering algorithm has been widely used in spectral processing due to its advantages of such as simple principles, powerful functions and only two parameters setting (the window size and the fitting order). But how to set the parameters of the S-G algorithm correctly to balance the filtering effect between insufficient denoising and excessive filtering is a big problem for its application. In the designed detection system, the spectral signal of the measured optical path is non-stationary signal, and the amplitudes of the noises and effective signals are time-varying. So the optimal window size and polynomial order are changing with the signal dynamics of large range. As a result, it’s difficult to achieve the optimal filtering effect through S-G filters with fixed parameters. With the adaptive hierarchical S-G smoothing filter algorithm proposed in this paper, the sum of the signal correlation coefficient and the first derivative of the signal from measured light path spectrum signal after S-G filtering layer by layer and the reference section set by the spectrum signal of the reference light path have been compared, and then the optimal parameters of each layer can be obtained adaptively. With the analysis on 10 groups of band noise spectrum of which the signal to noise ratios(SNR) are from 981 to 2977, the algorithm could effectively restore the concentration information carried by the band noise signals of the gas to be measured. Compared with the band noise spectrum, the maximum error of the absorption spectrum peak has dropped from 25152% to 5917%, and the maximum error of the integral absorbance has decreased from 18.1% to 3.9%. In the realized system, the adaptive algorithm has been used for the filtering processing of the measured optical path. The CO concentration emitted by motor vehicles of different models, displacement and oil product use has been monitored online in real time when they passed the system at idle and low speed(5 km·h-1).