Aiming to meet the requirements of edge-aware and edge-planning navigation, it is necessary to perform real-time headland detection to determine the operable area and effectively extract the tillage boundary within the operable region for dynamic planning of navigation paths. In response to challenges such as large headland detection errors caused by the coexistence of tilled and untilled zones in non-homogeneous field scenarios during stubble field tillage operations, and reduced accuracy of tillage boundary extraction due to environmental factors like lighting, a headland detection method was proposed based on the partitioned dispersion of the average grayscale values of image rows, and a tillage boundary detection method using pre-extracted boundaries to define a dynamic region of interest. The headland detection was performed by analyzing the grayscale variation trends in the color space. A single frame was partitioned to evaluate the horizontal distribution of average grayscale values, and an independent dynamic threshold was used to determine the presence of a headland. For tillage boundary navigation line extraction, coarse superpixels were firstly used for preliminary image segmentation to extract pseudo navigation lines and determine the region of interest. Then a four-direction bidirectional gradient adaptive weight total variation algorithm was applied for noise filtering and denoising. The region image was finely segmented by using a two-dimensional cross-entropy method. Finally, the Canny operator was used to extract edge feature points, and pre-screening was performed on boundary points to be fitted. The navigation line was then fitted by using the random sample consensus algorithm, ultimately achieving accurate navigation line detection. Experimental results showed that the proposed headland recognition method achieved a detection accuracy of 96.04%, with an average processing time of 11.17 ms/f. The navigation line extraction method yielded an average angular deviation of 1.31° from manually annotated navigation lines. At the median image height, the average horizontal pixel deviation and spatial deviation were 10.95 pixels and 32.04 mm, respectively. The average processing time for navigation line extraction was 86.65 ms/f. This demonstrated the method's capability for stable and effective navigation line extraction, providing a reference for autonomous rotary tillage operations in stubble fields.