<p>Accurate tomato localization in unstructured greenhouse environment is essential for the advancement of automated agricultural robots. Existing fruit localization algorithms, such as the Census transformation algorithm, often struggle with computational speed and accuracy. This paper proposes an optimized Census stereo matching algorithm tailored for such complex scenarios. By utilizing the average pixel value of the matching window center instead of the center pixel alone, we reduce noise interference and enhance matching accuracy. Furthermore, we restrict the matching region to the predicted tomato area and limit the stereo matching search range to the robotic arm’s workspace—these optimizations significantly reduce computational load and improve efficiency. Experimental results demonstrate that our enhanced algorithm achieves an average localization error of 6.87&#xa0;mm with a single image processing time of less than 60ms, representing a substantial improvement over traditional method. This work not only enhances the practical value of high efficiency automated agricultural harvesting but also presents opportunities for future research in fruit localization technologies.</p>