{"id":912,"date":"2017-08-28T13:54:01","date_gmt":"2017-08-28T11:54:01","guid":{"rendered":"http:\/\/www.calvert.ch\/maurice\/?p=912"},"modified":"2017-12-06T23:42:06","modified_gmt":"2017-12-06T22:42:06","slug":"ublox-c94-m8p-rtk-gps-review-performance-evaluation","status":"publish","type":"post","link":"https:\/\/www.calvert.ch\/maurice\/2017\/08\/28\/ublox-c94-m8p-rtk-gps-review-performance-evaluation\/","title":{"rendered":"UBlox C94-M8P RTK GPS review and performance evaluation"},"content":{"rendered":"

TLDR: Results<\/a> . Lessons learned<\/a> . Summary & Conclusions<\/a> . Quick-start guide<\/a><\/p>\n

I am building an autonomous lawnmower which needs to know its position precisely \u2013 in the order of centimetres. Standard GPS is an order of magnitude too coarse and until recently, all flavours of differential GPS<\/a> were very expensive. This changed last year when UBlox released the C94-M8P<\/a> with a price tag of \u20ac359 for a complete solution. The announcement said:<\/p>\n

The NEO-M8P module series introduces the concept of a \u201cRover\u201d and a \u201cBase Station\u201d. By using a correction data stream from the Base Station, the Rover can output its relative position with stunning cm-level accuracy in good environments.<\/p><\/blockquote>\n

An old dog, I was wary of \u201cin good environments\u201d. That RTK<\/a> is capable of centimeter accuracy when stationary and with a horizon-to-horizon view of the sky, I have no doubt; but what about in a less GPS-friendly environments? This article is the result of real-life testing in my garden, cluttered by surrounding buildings and vegetation..<\/p>\n

Objectives<\/h1>\n

My aim was to establish whether or not the UBlox solution would be sufficiently accurate for my mower,
\nin particular how would it perform in less-than optimal conditions, so I devised a series of experiments, which I have named for the purposes of this discussion.<\/p>\n

    \n
  1. Open<\/em>. Stationary for 24 hours in the middle of the lawn, with a decent view of the sky. A base-line test to determine the system performance.<\/li>\n
  2. Sky270<\/em>. The receiver is next to the corner of a building, 90\u00b0 of sky are obscured.<\/li>\n
  3. Sky180<\/em>. The receiver is close to a wall of a building, 180\u00b0 of sky are obscured.<\/li>\n
  4. Sky090<\/em>. The receiver is in a corner of a building, 270\u00b0 of sky are obscured.<\/li>\n
  5. Canyon<\/em>. The receiver is between two buildings with a heavy tree canopy, most of the sky is obscured.<\/li>\n
  6. Mobile<\/em>. The receiver is moved around the garden, stopping for 10 seconds at a known location in each of the previous tests.<\/li>\n<\/ol>\n

    Tests 2-6 are performed for an hour. As described later, I discovered that the receiver needs 25-30 minutes to settle on a solution and these data were removed.<\/p>\n

    All the tests are made with the antenna 54cm from the ground.<\/p>\n

    Test setup<\/h1>\n

    I built a rooftop base station<\/p>\n

    \"\"<\/a><\/p>\n

    powered by a solar cell with an MPP<\/a> charger, the DC1621A<\/a> evaluation board based on the LTM8062<\/a> chip from Linear Technology. As I had no idea how much power the C94-M8P would need \u2013 after all it has a UHF transmitter \u2013 I opted for a 10Ah 2S Lipo. On the left we have the DC1621A, on the right the UBlox board and the battery below. There\u2019s a cell-balancer and packets of silica gel tucked behind the DC1621A. After verifying that every was working correctly, I mounted it on my roof:<\/p>\n

    \"\"<\/a><\/p>\n

    My rover was simply a low table in the garden, with the antenna 54cm off the ground.<\/p>\n

    The rover’s signature is<\/p>\n

    \"\"<\/a><\/p>\n

    In what follows, I am focussed solely on horizontal position \u2013 the X, Y values; I ignore altitude as it is only relevant for surveying and aviation.<\/p>\n

    All my calculations were performed in 64-bit floating point which gives ~15 decimal digits of precision.<\/p>\n

    The raw GPS data from the receiver contains latitude\/longitude values. To convert these into distances I use Vincenty\u2019s formula<\/a>. I validated my implementation against 500\u2019000 known distances<\/a>, which are precise to 0.1\u00b5 metre.<\/p>\n

    Outputs<\/h1>\n

    Each test produces 5 results.<\/p>\n

    1. Summary statistics<\/h3>\n

    \"\"<\/a><\/p>\n

    Quality<\/strong> (the NMEA term) is the type of Fix that the GPS currently has. Possible values are:<\/p>\n

      \n
    1. Autonomous GNSS<\/li>\n
    2. Differental GNSS<\/li>\n
    3. 3D fix<\/li>\n
    4. RTK Fixed<\/li>\n
    5. RTK Float<\/li>\n<\/ol>\n

      It is worth noting that the ‘best’ fix is 4.<\/p>\n

      SVs<\/strong> is the number of satellites in view. I configured the rover UBX-CFG-NMEA to high precision mode (to output 9 digits of precision on latitude and longitude) and as a result considering mode is disabled. It is unclear whether satellites-in-view is those seen or those actually used.<\/p>\n

      hDop<\/strong> is the horizontal dilution of precision created by the constellation geometry.<\/p>\n

      ErrorLat<\/strong> is the distance in millimetres from the average centre of all the readings. Similarly ErrorLon.<\/p>\n

      ErrorLatLon<\/strong> is Euclidean distance from the measured point and the averaged centre of all the readings.<\/p>\n

      The ErrorX values are calculated using Vincenty’s Formul\u00e6.<\/p>\n

      2. Error statistics<\/h3>\n

      \"\"<\/a><\/p>\n

      Limit<\/strong> is simply a reminder of the percentages associated with the number of standard deviations.<\/p>\n

      \u03c3<\/strong> is the standard sigma calculation, the square root of the variance.<\/p>\n

      Out<\/strong> is the actual number of measurements that exceed the CEP below.<\/p>\n

      CEP<\/strong> is obtained by finding the number of samples ‘Out’ which exceed the limit %.<\/p>\n

      The difference between \u03c3<\/strong> (a statistical calculation) and CEP<\/strong> (the true value for the data set in question) is readily explained: The statistical \u03c3 is valid only when the data are normally distributed (a Gaussian). Even over 24 hours, GPS readings are not Gaussian because the average is not stationary. Disraeli’s quip<\/a> remains as true as ever.<\/p>\n

      3. Deviation map<\/h3>\n

      An X-Y plot of ErrorLat horizontally against ErrorLon vertically:<\/p>\n

      \"\"<\/a><\/p>\n

      A point’s colour indicates its age as hue corrected for the human eye. First blue, then green, yellow, orange and finally red.<\/p>\n

      4. ErrorLatLon over time<\/h3>\n

      The Euclidean distance from the average centre to the measured point, in mm, over time:<\/p>\n

      \"\"<\/a><\/p>\n

      5. Histogram<\/h3>\n

      The number of errors at each bucket of the Euclidean distance from the average centre:<\/p>\n

      \"\"<\/a><\/p>\n

       <\/p>\n

      Test results<\/h1>\n

      Open<\/em> test<\/h2>\n

      Sky visibility is typical for a residential area. To the north and west a single-storey building<\/p>\n

      \"\"<\/a><\/p>\n

      The receiver, antenna and PC are on a low table in the foreground.<\/p>\n

      To the east, a 4-storey building (actually 3-storey but on higher ground)<\/p>\n

      \"\"<\/a><\/p>\n

      To the south and west, mature oak trees and 2-storey houses<\/p>\n

      \"\"<\/a> \"\"<\/a><\/p>\n

      The receiver is in Fixed mode 71% of the time.<\/p>\n

      \"\"<\/a><\/p>\n

      \"\"<\/a><\/p>\n

      The 3\u03c3 CEP of 712 mm is surprisingly high and this is visible both in the deviation map:<\/p>\n

      \"\"<\/a><\/p>\n

      and in the histogram (I have added log(Count) overlayed on Count to show the long tail more clearly):<\/p>\n

      \"\"<\/a><\/p>\n

      The problem occured mid-morning and early afternoon:<\/p>\n

      \"\"<\/a><\/p>\n

      The cause appeared to be due to receiver occasionally losing the RTK solution. To analyse this I zoomed in on the first occurrence and plotted the absolute error along with the Fix. 3D=2, Fixed=4 and Float=5. To highlight the differences, Fixed errors are green, Float errors are blue and 3D errors are red. Notice that the vertical scale is logarithmic:<\/p>\n

      \"\"<\/a><\/p>\n

      At 10:28:18 and 10:32:52 the receiver switches for one measurement from Float to 3D. The same thing happens at 11:21:18:<\/p>\n

      \"\"<\/a><\/p>\n

      and in both cases takes some 20 minutes before settling down.<\/p>\n

      The replay in ucentre shows nothing untoward at the instant of the loss:<\/p>\n

      \"\"<\/a><\/p>\n

      I note that 20 minutes is the time that it takes the receiver to lock on cold start (details in Lessons Learned below) and postulate that momentary interference can cause a complete loss of the solution.<\/p>\n

      There were 4 such events in the 24-hour test. 4 in 86’400 = 1 in 21’600; a frequency less than 4\u03c3, so I sought an explanation.<\/p>\n

      On the weekend of the tests there was an airshow nearby with several passes of military aircraft in formation. One can reasonably assume that 6 fighters at low altitude produce vast quantities of electromagnetic radiation and I imagine that military RF applications care little about the interference that they generate. Furthermore, the dropouts only occurred during the day, when the aircraft were flying, so I decided that they were the most likely cause. Consequently I decided to use a subset of the data with these events excluded. I also excluded the noisier data as of 04:32 in the morning, on ‘benefit of the doubt’ reasoning, for want of an explanation. The results then become:<\/p>\n

      \"\"<\/a><\/p>\n

      \"\"<\/a>\u00a0\"\"<\/a>\u00a0\u00a0\"\"<\/a><\/p>\n

      Which is what I expected and much more reasonable.<\/p>\n

      Sky270<\/em> test<\/h2>\n

      The receiver is against the NW corner of the house, 90\u00b0 of obscured sky.<\/p>\n

      \"\"<\/a><\/p>\n

      The receiver is slightly more in Float rather than Fixed (4.63 average quality). The 3\u03c3 CEP is 54 mm.<\/p>\n

      \"\"<\/a><\/p>\n

      \"\"<\/a> \"\"<\/a> \"\"<\/a><\/p>\n

      There is significantly more longitude error than latitude.<\/p>\n

      Sky180<\/em> test<\/h2>\n

      The receiver is near the NW wall of the house, half the sky is obscured.<\/p>\n

      The receiver is continuously in Fixed mode. The 3\u03c3 CEP is 43 mm.<\/p>\n

      \"\"<\/a><\/p>\n

      \"\" <\/a>\"\"<\/a> \"\"<\/a><\/p>\n

      The histogram has longer tails due to the less favourable constellation geometry (practically no SVs to the SE):<\/p>\n

      \"\"<\/a><\/p>\n

      The CEP is better than the Sky270 test, where only 90\u00b0 of sky were obstructed. I attribute this to the fact that due to the lay of the land, the angle of the apex of the house is much larger when viewed from a corner. Additionally the sky view away from the house in this position is a little more cluttered.<\/p>\n

      Sky90<\/em> test<\/h2>\n

      The receiver is in a corner of a building, facing SW, 270\u00b0 of sky are obscured.<\/p>\n

      The receiver is in Fixed mode for the whole hour. The 3\u03c3 CEP is 185 mm.<\/p>\n

      \"\"<\/a><\/p>\n

      \"\"\u00a0<\/a>\"\"<\/a>\u00a0<\/a>\"\"<\/a><\/p>\n

      The deviation shape clearly reflects the SW direction of the corner.<\/p>\n

      Canyon<\/em> test<\/h2>\n

      The receiver is in a NE-SW canyon between two houses with heavy canopy:<\/p>\n

      \"\"<\/a><\/p>\n

      The wide-angle lens gives the impression that there is a wide view towards the camera. This is not the case, the two houses’ walls are almost parallel, some 12 metres apart and the canopy almost reaches to above the camera.<\/p>\n

      The receiver never achieves Fixed mode. The 3\u03c3 CEP is 135 mm.<\/p>\n

      \"\"<\/a><\/p>\n

      \"\" <\/a>\"\"<\/a> \"\"<\/a><\/p>\n

      The error and histogram are noisier and the canyon effect is clear in the deviation map, where the longitude errors are double the latitude errors, again due the the constellation geometry (few SVs NE and SW):<\/p>\n

      \"\"<\/a><\/p>\n

      Mobile<\/em> test<\/h2>\n

      I walked around the garden, pausing 10 seconds at each of the SKYNNN<\/em> points to hold the GPS at a marked position.<\/p>\n

      \"\"<\/a><\/p>\n

      The red points are the 9 nearest points to the average of each location.<\/p>\n

      The receiver was in Fixed mode 7% of the time.<\/p>\n

      The average position error when moving is 19 mm.<\/p>\n

      \"\"<\/a><\/p>\n

      Lessons learned<\/h1>\n

      Here are some tips which might be useful to those envisaging the C98-M8P solution.<\/p>\n

      Status<\/h3>\n

      The status bar at the bottom-right of U-Centre is very helpful:<\/p>\n

      \"\"<\/a><\/p>\n

      From left to right:<\/p>\n