The cigarette mainstream smoke represents the main gas that is ingested by the human body when the cigarette is burned. The reduction of scorch and other hazardous components has become an issue of great concern to the whole society. Among various components in cigarette mainstream smoke, crotonaldehyde has become one of the seven main harmful indicators in cigarettes prescribed by the state due to its strong genotoxicity. Traditional analytical methods for crotonaldehyde usually rely on high-performance liquid chromatography and other laboratory methods, which requires complex sample pretreatment procedures. This makes it difficult to measure crotonaldehyde in real-time to evaluate its effects on health. In order to monitor the crotonaldehyde content in cigarette mainstream smoke efficiently, a Fourier Transform Infrared Spectrometer (FTIR) system was set up to a couple with a smoking machine. In this system, an innovative oversampling data driven spectral analysis (ODDSA) method was developed to accurately extract the spectral features of crotonaldehyde from the complex and fluctuating spectra of cigarette mainstream smoke. The ODDSA method started with experimental design and used the idea of random design to simulate the distribution range of actual cigarette samples, which constructed a good data structure to guide further data mining. Thereafter, the high-density wavelet transform (HDWT) was innovatively used to process the IR spectra, which enabled oversampling in time/frequency dual-domains to improve the spectral resolution. This would definitely suppress the effects of other matrix components on the analysis of crotonaldehyde. Finally, the strategy of modified competitive adaptive reweighted sampling was developed to accurately extract the interseting features from the multiple redundant HDWT coefficients, which was used to construct a qualified calibration model for the analysis of crotonaldehyde. In the experiment, 15 typical commercial cigarette brands were collected, in which 8 samples of each brand were prepared to collect their IR spectra of mainstream smoke. Thereafter, 25 samples were randomly selected to validate the performance of ODDSA. The calculation results showed that the regression coefficient of the test set was 0.971, and the relative root means square error is 5.5%. The satisfactory results indicate that the ODDSA is capable of on-line analysis of crotonaldehyde in cigarette mainstream smoke, which may well extend to on-line monitoring of other components in second-hand environmental smoke. This would provide a novel tool for the evaluation of cigarette effects on health.