Three attitude determination algorithms are compared in this paper. The three methods are the complementary filter, a quaternion-based Kalman filter and a quaternion-based gradient decent algorithm. To investigate and compare their performance, experiments were conducted and the results analysed. This paper shows that the complementary filter requires the least computational power; quaternion based gradient decent algorithm has the best noise filtering ability; quaternion based gradient algorithm has the highest accuracy. Since each algorithm makes use of the quaternion, the quaternion to Euler angle singularity property must be investigated. Experiments conducted show that when Y-rotation approach the singularity position (±90°), the X-rotation drifts away from the reference input. This paper proposes the use of a set of imaginary input which replaces the original input during Y-rotation approaching the singularity position.