Stringent requirements on the accuracy of attitude determi- nation are currently a major challenge for strapdown INS/GPS integration, which is at the core of self-contained airborne remote sensing and mapping systems. This paper reviews the error models for INS/GPS integration and focuses i n detail on designingfiltering methodsfor improvingattitudeaccuracyin the bandwidth in which an inertial system does not benefit from frequent GPS position/velocity updates. Several fltering methods are designed based on the spec- tral analysis of the raw inertial signal in a dynamic environ- ment. These include a spectral technique for dither spike removal and a class of low-pass finite-impulse-response (FIR] filters operatinginforward/backward mannerfor achieving zero phase distortion. The orientation performance of the whole system with different filters i s evaluated b y comparing it to the "true" attitude information provided by a photo- grammetric block adjustment. Results show clearly that the choice of an appropriatefilter is decisivefor attitude accuracy. Overall, the INS/GPS integration combined with the most suitable pre-filtering method agrees with the external orien- tationreferenceto0.005"(19')RMS overthewholetestperiod, while the flight-line consistency is typically 0.003" (1U')RMS. The best filter in the comparison has an i?MS seven times smaller than the Butterworth filter which i s frequently applied i n the industrial designs of INS.