First agricultural revolution – ~12000BC Second agricultural revolution – 18th&19th Century Current situation • High dependency on fossil fuels What’s next? • High dependency on chemicals Sustainable Agriculture • Advanced genetics for increased crop yield • Fast growing human population
Weed control is a fundamental operation for any crop to maintain high yields 90% is performed chemically Large machinery • Oil consumption CO 2 emission • Weeding alternatives • Mechanical weeding (Hoeing machine) – Not efficient • Manual weeding – 10 times more expensive
Untargeted application… • >90% wasted → high costs • Toxic residues on soil, water, crop → health & environment impact • Damage on crops → 5 -10 % yield losses in average …of selective herbicides • New molecules expensive to develop ($350M 1 ) • Increased legal pressure to ban molecules & reduce quantity • Development of herbicide-resistant weeds (+10%/year 2 ) 1 Phillips McDougall, 2015 2 “Global Increase in Unique Resitant Cases ”, Dr Ian Heap, Weedscience.org, 2016
Autonomous weeding / navigation • Minimal human intervention Solar powered • Sustainable energy & Power autonomy AI for weed detection Robotic arms for precision spraying • Same efficiency up to 40% cheaper • No chemicals on crops • Up to 20 times less herbicide • Allows cheaper & ecological molecules • Reduces herbicide-resistant weeds problem
Solar panels, 350 W max Delta robotic arms Electric motors, 0.5 m/s speed GPS & IMU for navigation RGB Camera, 5 MP @10 fps Jetson TK1 (TX2) for CV & ML Weed detection @ >1 fps
Row Finder Pipeline Plants Row Image Downscaling Segmentation Finder Camera Settings Adjustment & Weed Detect Pipeline Image Acquisition Plants Feature Classification Segmentation Extraction Send weed positions Map weed positions Post-processing & sizes to arms to robot coord.
Row Finder Pipeline Plants Row Image Downscaling Segmentation Finder Camera Settings Adjustment & Weed Detect Pipeline Image Acquisition Plants Feature Classification Segmentation Extraction Send weed positions Map weed positions Post-processing & sizes to arms to robot coord.
Plant Pre-Segmentation JETSON TK1 JETSON TX2 CPU GPU CPU GPU 424.6 ± 18.1 55.6 ±1.86 130.9 ± 27.9 36.7 ± 10.8 Color Space Conversion 107.2 ± 2.6 0.41 ± 1.21 79.9 ± 15.7 0.21 ± 0.71 Color Normalization 208.9 ± 6.1 41.3 ± 8.9 196.7 ± 6.4 14.7 ± 4.92 Gaussian Smoothing 881.3 ± 20 98.45 ± 9.1 517.3 ± 42.4 51.8 ± 14.1 Overall (Mean Computation Times over 90 images in ms)
PROTOTYPE 1 PROTOTYPE 2 HW: JETSON TK1 HW: JETSON TX2 Hand-crafted Features DNN AdaBoost Classifier DNN
PROTOTYPE 1 PROTOTYPE 2 HW: JETSON TK1 HW: JETSON TX2 Hand-crafted Features DNN AdaBoost Classifier DNN Set 1 – Seen field OA = 92% F1=0.87 OA = 92.2% F1=0.867 (w.r. 33.9%) Prec. = 97.2% Rec. = 78.3% Prec. = 96.3% Rec. = 78.8% Set 2 – Seen field OA = 97.8% F1=0.77 OA = 97.8% F1=0.78 (w.r. 5.6%) Prec. = 94.4% Rec. = 65.6% Prec. = 87.3% Rec. = 70.5% OA = Overall Binary Accuracy, F1 = Mean F-Score Prec. = Weed detection precision, Rec. = Weed detection recall w.r. = Weed / (Weed + Crop) pixels ratio
Effects of Shadow on Color & Visibility
Illumination Changes due to daytime & weather 10am 6pm
Soil & Crop Variation Field 1 Field 2
PROTOTYPE 1 PROTOTYPE 2 HW: JETSON TK1 HW: JETSON TX2 Hand-crafted Features DNN AdaBoost Classifier DNN Mixed Set – Unseen Fields OA = 79.4% F1=0.531 OA = 92.5% F1=0.863 (w.r. : 23.4%) Prec. = 56.7% Rec. = 49.9% Prec. = 86.5% Rec. = 86.1% OA = Overall Binary Accuracy, F1 = Mean F-Score Prec. = Weed detection precision, Rec. = Weed detection recall w.r. = Weed / (Weed + Crop) pixels ratio
JETSON TK1 – JETSON TX2 – Jetson TX2 – Feat. Extraction + AdaBoost Feat. Extraction + Adaboost Deep Learning (CNN) 987.1 ± 523.7 ms 744 ± 503.4 ms 634.5±393.5 ms After Optimization 780 ± 447.6 ms Full Pipeline Full Pipeline Full Pipeline 977.4 ± 511.9 ms 833.8 ± 397.6 ms 879 ± 457.5 ms (Mean Computation Time over 20 images)
Real-life embedded GPU application that can improve food production First completely autonomous weeding robot NVIDIA Jetson TX2 opens new frontiers for embedded platforms Deep Learning is powerful, but it is no magic wand Challenging to obtain images covering all situations
Contact: An ıl Yüce anil.yuce@ecorobotix.com
Recommend
More recommend
