Difference between revisions of "Robotic Yield Estimation of Apple Orchards"

From RSN
Jump to: navigation, search
(Performance Evaluation of Yield Mapping:)
Line 27: Line 27:
 
|}
 
|}
  
In the next video, we show the performance of our counting algorithm from the sunny side of a row. The conservative and non-conservative tags are based on a supervised detection procedure. The detection process labels the pixels which has a large probability of belonging to apples as conservative, and ones with lesser probabilities as non-conservative. Median accuracy for counting with respect to visible Apples from both sides is 93.86% and accuracy with respect to actual is 73.71%. Mean accuracy with respect to visible apples from both sides is 88.20%, and accuracy with respect to actual is 69.25%
+
In the next video, we show the performance of our counting algorithm from the sunny side of a row. The conservative and non-conservative tags are based on a supervised detection procedure. The detection process labels the pixels which has a large probability of belonging to apples as conservative, and ones with lesser probabilities as non-conservative. Median accuracy for counting with respect to visible apples from both sides is 93.86%.
As the conservative and non-conservative detection performance was similar for this particular data set, the non-conservative performance results were exactly same.
 
  
 
{| cellpadding="0" cellspacing="10" align="center"
 
{| cellpadding="0" cellspacing="10" align="center"
Line 36: Line 35:
  
  
Next we show the performance for the shady side of the same row. Median accuracy for counting with respect to visible Apples from both sides is 80.18% and accuracy with respect to actual is 62.9%. Mean accuracy with respect to visible apples from both sides is 76.88%, and accuracy with respect to actual is 60.3%. Again, the conservative and non-conservative detection performance was similar for this particular data set, the non-conservative performance results were exactly same.
+
Next we show the performance for the shady side of the same row. Median accuracy for counting with respect to visible Apples from both sides is 80.18%.
  
 
{| cellpadding="0" cellspacing="10" align="center"
 
{| cellpadding="0" cellspacing="10" align="center"

Revision as of 18:31, May 2, 2017

Overview of our yield mapping work

Performance Evaluation of Yield Mapping:

We developed algorithms for detecting apples, counting them and estimating their diameters from a video sequence.

For validation of our algorithms we hand labelled the apples both in the real world and in images. This allowed us to obtain ground truth to evaluate our algorithms.


Hand Labelled Apples

Propagation of Hand Labels Using Homography

In this video we show how our detection method is performing. The blue boxes are detection by the algorithm and the green boxes are the manually labelled ones. We detect 97.861% of the hand labelled apples which we call, “Hit Rate”.

Hit Rate

In the next video, we show the performance of our counting algorithm from the sunny side of a row. The conservative and non-conservative tags are based on a supervised detection procedure. The detection process labels the pixels which has a large probability of belonging to apples as conservative, and ones with lesser probabilities as non-conservative. Median accuracy for counting with respect to visible apples from both sides is 93.86%.

Counting Performance for Sunny Side of a Row


Next we show the performance for the shady side of the same row. Median accuracy for counting with respect to visible Apples from both sides is 80.18%.

Counting Performance for Shady Side of the Same Row


This project is funded by USDA NIFA MIN-98-G02.

Related Publications

2016
2Z. Li, V. Isler
Large Scale Image Mosaic Construction for Agricultural Applications
IEEE Robotics and Automation Letters, 1(1): 295-302, 2016.
1N. Stefas, H. Bayram, V. Isler
Vision-Based UAV Navigation in Orchards
In 5th IFAC Conference on Sensing, Control and Automation Technologies for Agriculture, 2016.