A rapid estimation of intra-row weed density using an integrated CRM, BTSORT and HSV model across entire video stream of chilli crop canopies

Abstract

Intra-row weeds have a significant impact on crop yield and health, competing with crops for nutrients, water, and sunlight. Traditional uniform herbicide applications in weed management not only risk harming the environment but can also compromise crop health. Precision spraying technology, guided by machine vision, offers a solution by accurately identifying and targeting weeds, thereby reducing overall herbicide use. Inter-row weed segmentation can be done easily with simple thresholding, but intra-row region weed cannot be segmented with simple thresholding due to many similarities between the intra-row weeds and plants. So, in this study, a novel methodology is introduced to dynamically estimate intra-row weed density for the entire crop row of chilli by integrating ByteTrack Simple Online and Real Time Tracker (BTSORT) with YOLOv7 crop recognition model to track the plant and Hue-Saturation-Value (HSV) color model with simple thresholding to segment weeds between tracks to avoid repetitive intra-row weed density estimation. The weed density between the plants in these regions is calculated and categorized into low, medium, and high levels based on the number of weed pixels in the intra-row region. The YOLOv7 Crop recognition model recognized the chilli plants with achieved a precision of 0.92 and a recall of 0.94 at 47.39 FPS. The BTSORT with YOLOv7 crop recognition model on a test video dataset performed well with MOTA and MOTP of 0.85 and 0.81, respectively. The developed dynamic intra-row weed density estimation method classifies it with an overall accuracy of 0.87. Additionally, the system processed 1280x720 frames 1.38 times faster than 1920x1080 frames, enabling efficient real-time intra-row weed density estimation across full crop rows.