Objective  The precise location algorithm of lawn weeds based on a fusion algorithm was explored to provide theoretical basis and technical support for lawn weed automatic identification and removal robots to carry out weeding operations.

Method A multi-method fusion algorithm based on semantic segmentation for lawn location was proposed. First, the PSPNet network segmented the lawn and non-lawn contours. Secondly, for the segmented non-lawn contours, the interest area was extracted, excluding the non-weed contours and retaining the weed contours. Then, Zhang-Suen thinning algorithm extracted weed contour skeleton lines and obtained the number and coordinates of skeleton intersections. Finally, The fusion algorithm selected different positioning strategies according to the number of intersecting points to achieve accurate positioning of weed roots.

Result The root-mean-square error between the location coordinates of weeds by fusion algorithm and the actual root center coordinates of weeds was 83.17 Px, and reduced by 14% compared with the method of mean centroid, reduced by 22% compared with the method of a minimum circumscribed circle. Converted to the actual scenario, the root-mean-square error between the location coordinates of weeds by fusion algorithm and the actual root center coordinates of weeds was 12.48 mm, which was within the acceptable error range. The fusion algorithm can improve the accuracy of weed root center location, and reduce the error of weed location.

Conclusion The location method of lawn weeds based on semantic segmentation and fusion algorithm can improve the localization accuracy of weed root centers, reduce the positioning error by a single method, and provide technical support for automatic identification and removal of lawn weeds.