There are several methods to get individualized HRTFs such as measurements, syntheses, or selection. The study first aims to select the fittest HRTFs in an existing database for listeners. The idea is developed based on the connection between anthropometric parameters and auditory localisation. By means of machine learning, the neural network designed for HRTF ranking produces a reliable prediction. The new approach has been verified by the anthropometry and listening perception of 24 subjects. The final selected dataset is used to synthesize a virtual audio scene. Two binaural ambisonic decoders are proposed to overcome the dilemma of improving sound localisation or enhancing audio quality. The first one is the equalization decoding, equalizing the root mean square (RMS) power in each 1/3 octave frequency bank, especially compensating the low-pass filtered components in a high-density speaker array. Therefore, the energy distribution of the treated signal is nearly uniform. The other proposed method is split-band decoding, selecting and then mixing the better reconstructed frequency components from different speaker arrays. Through several experiments and listening tests, there are no click noises in the real-time system, when the virtual auditory space is rapidly rotated. The split-band decoder presents comparable performance to a pure HRTF decoder.