To improve the recognition ability of a pyroelectric infrared (PIR) detector for different infrared radiation sources, a method for human and non-human recognition based on Canonical Correlation Analysis (CCA) was proposed. Firstly, the frequency spectrum and wavelet packet entropy were extracted as features,and the spectrum was divided into sub-patterns. Then, each sub-pattern and wavelet packet entropy were fused with CCA method, and the fused feature was employed as classification information. By this way, the feature fusion was realized and the redundant information among the features was also eliminated. Finally, the recognition results were obtained by a majority voting method. As a special case of the sub-pattern fusion, the classification abilities of the features fused with their own sub-pattern were also studied in the paper. Experimental results show when the frequency is divided into 5 sub-patterns, the recognition rate can reach 95.2%, which is higher 2.7% than that of only fusing the frequency and the wavelet packet entropy.Moreover, the recognition rate of wavelet packet entropy fused with its own sub-pattern is 90.7% , which is higher 2.3% than that of wavelet packet entropy.