In this paper, we propose a high-speed part recognition method that can be practically used even in cases of drastic illumination variation. Our method selects a small number of distinctive and stable pixels in template images by analyzing spatial and temporal co-occurrence probabilities; these pixels are used in the matching process. Reduced pixels enable matching to be speeded up and stable pixels guarantee robustness against varying degrees of illumination. In addition, the proposed method optimizes the pixel selection criteria dynamically so they can be utilized under long-term illumination changes. It has been demonstrated that our system achieves a 97.6% recognition success rate through an experiment with 24120 test images that included drastic illumination variation, such as that occurring when the sun sets on a clear day. We proved that recognition success rate of the proposed method is approximately 30% higher than our previous work (CPTM) which using only of spatial co-occurrence.