class _OdometerShape(object):
# A 0.1% cone - Coefficient to defavor points further away on a shape
# when doing stop snapping to shape. The summit angle of the offset cone
# is 2 * arctan(K).
K = 0.001
def __init__(self, shape):
self._shape = shape
self._cache = _CacheEntry(None)
self._cache_hit = 0
self._cache_miss = 0
if all(pt.shape_dist_traveled != -999999 for pt in shape.points):
self._xdist = ContinousPiecewiseLinearFunc()
else:
self._xdist = None
# Normalize the shape here:
# 1) dist_traveled to meters
# 2) pt_seq to contiguous numbering from 0
ptseq = 0
distance_meters = 0.0
last_pt = None
shape.points.sort(key=lambda p: p.shape_pt_sequence)
for pt in shape.points:
if last_pt is not None:
# Note: we do not use distance cache, as most probably
# many of the points will be different from each other.
distance_meters += orthodromic_distance(last_pt, pt)
last_pt = pt
pt.shape_pt_sequence = ptseq
old_distance = pt.shape_dist_traveled
pt.shape_dist_traveled = distance_meters
# Remember the distance mapping for stop times
if self._xdist:
self._xdist.append(old_distance, pt.shape_dist_traveled)
ptseq += 1
def reset(self):
self._distance = 0
self._istart = 0
self._cache_cursor = self._cache
def dist_traveled(self, stop, old_dist_traveled):
if old_dist_traveled and self._xdist:
# Case 1: we have in the original data shape_dist_traveled
# We need to remap from the old scale to the new meter scale
return self._xdist.interpolate(old_dist_traveled)
else:
# Case 2: we do not have original shape_dist_traveled
# We need to determine ourselves where in the shape we lie
# TODO Implement a cache, this can be slow for lots of trips
# and the result is the same for the same pattern
# Check the cache first
cache_entry = self._cache_cursor.next_entry(stop)
if cache_entry is not None:
self._cache_cursor = cache_entry
self._cache_hit += 1
return cache_entry.distance()
min_dist = 1e20
best_i = self._istart
best_dist = 0
for i in range(self._istart, len(self._shape.points) - 1):
a = self._shape.points[i]
b = self._shape.points[i + 1]
dist, pdist = orthodromic_seg_distance(stop, a, b)
newdist = a.shape_dist_traveled + pdist
howfar = newdist - self._distance
# Add a slight "cone" offset. There are pathological
# cases with backtracking shapes where the best distance
# is slightly better way further (for eg 0.01m) than at
# the starting point (for eg 0.02m). In that case we should
# obviously keep the first point instead of moving too fast
# to the shape end. That offset should help for some cases.
dist += howfar * self.K
if dist < min_dist:
min_dist = dist
best_i = i
best_dist = newdist
if best_dist > self._distance:
self._distance = best_dist
else:
delta = self._distance - best_dist
if delta > 10:
# This is harmless if the backtracking distance is small.
# We have lots of false positive (<<1m) due to rounding errors.
logger.warn(
"Backtracking of %f m detected in shape %s for stop %s (%s) (%f,%f) at distance %f < %f m on segment #[%d-%d]"
% (delta, self._shape.shape_id, stop.stop_id,
stop.stop_name, stop.stop_lat, stop.stop_lon,
best_dist, self._distance, best_i, best_i + 1))
self._istart = best_i
self._cache_miss += 1
self._cache_cursor = self._cache_cursor.insert(
stop, self._distance)
return self._distance
def _debug_cache(self):
logger.debug("Shape %s: Cache hit: %d, misses: %d" %
(self._shape.shape_id, self._cache_hit, self._cache_miss))
class _OdometerShape(object):
# A 0.1% cone - Coefficient to defavor points further away on a shape
# when doing stop snapping to shape. The summit angle of the offset cone
# is 2 * arctan(K).
K = 0.001
def __init__(self, shape):
self._shape = shape
self._cache = _CacheEntry(None)
self._cache_hit = 0
self._cache_miss = 0
if all(pt.shape_dist_traveled != -999999 for pt in shape.points):
self._xdist = ContinousPiecewiseLinearFunc()
else:
self._xdist = None
# Normalize the shape here:
# 1) dist_traveled to meters
# 2) pt_seq to contiguous numbering from 0
ptseq = 0
distance_meters = 0.0
last_pt = None
shape.points.sort(key=lambda p: p.shape_pt_sequence)
for pt in shape.points:
if last_pt is not None:
# Note: we do not use distance cache, as most probably
# many of the points will be different from each other.
distance_meters += orthodromic_distance(last_pt, pt)
last_pt = pt
pt.shape_pt_sequence = ptseq
old_distance = pt.shape_dist_traveled
pt.shape_dist_traveled = distance_meters
# Remember the distance mapping for stop times
if self._xdist:
self._xdist.append(old_distance, pt.shape_dist_traveled)
ptseq += 1
def reset(self):
self._distance = 0
self._istart = 0
self._cache_cursor = self._cache
def dist_traveled(self, stop, old_dist_traveled):
if old_dist_traveled and self._xdist:
# Case 1: we have in the original data shape_dist_traveled
# We need to remap from the old scale to the new meter scale
return self._xdist.interpolate(old_dist_traveled)
else:
# Case 2: we do not have original shape_dist_traveled
# We need to determine ourselves where in the shape we lie
# TODO Implement a cache, this can be slow for lots of trips
# and the result is the same for the same pattern
# Check the cache first
cache_entry = self._cache_cursor.next_entry(stop)
if cache_entry is not None:
self._cache_cursor = cache_entry
self._cache_hit += 1
return cache_entry.distance()
min_dist = 1e20
best_i = self._istart
best_dist = 0
for i in range(self._istart, len(self._shape.points) - 1):
a = self._shape.points[i]
b = self._shape.points[i+1]
dist, pdist = orthodromic_seg_distance(stop, a, b)
newdist = a.shape_dist_traveled + pdist
howfar = newdist - self._distance
# Add a slight "cone" offset. There are pathological
# cases with backtracking shapes where the best distance
# is slightly better way further (for eg 0.01m) than at
# the starting point (for eg 0.02m). In that case we should
# obviously keep the first point instead of moving too fast
# to the shape end. That offset should help for some cases.
dist += howfar * self.K
if dist < min_dist:
min_dist = dist
best_i = i
best_dist = newdist
if best_dist > self._distance:
self._distance = best_dist
else:
delta = self._distance - best_dist
if delta > 10:
# This is harmless if the backtracking distance is small.
# We have lots of false positive (<<1m) due to rounding errors.
logger.warn("Backtracking of %f m detected in shape %s for stop %s (%s) (%f,%f) at distance %f < %f m on segment #[%d-%d]" % (
delta, self._shape.shape_id, stop.stop_id, stop.stop_name, stop.stop_lat, stop.stop_lon, best_dist, self._distance, best_i, best_i+1))
self._istart = best_i
self._cache_miss += 1
self._cache_cursor = self._cache_cursor.insert(stop, self._distance)
return self._distance
def _debug_cache(self):
logger.debug("Shape %s: Cache hit: %d, misses: %d" % (self._shape.shape_id, self._cache_hit, self._cache_miss))
def test_piecewise_linear_interpolation(self):
# Empty function
f = ContinousPiecewiseLinearFunc()
catched = False
try:
self.assertAlmostEqual(f.interpolate(0), 0, 3)
except:
catched = True
self.assertTrue(catched)
# Single point function
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(-1), 0, 6)
self.assertAlmostEqual(f.interpolate(1), 0, 6)
# A simple well-behaved function
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
f.append(1, 100)
self.assertAlmostEqual(f.interpolate(-0.000000001), 0, 6)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(0.000000001), 0, 6)
self.assertAlmostEqual(f.interpolate(0.5), 50.0, 6)
self.assertAlmostEqual(f.interpolate(0.9999999999), 100, 6)
self.assertAlmostEqual(f.interpolate(1.0), 100, 6)
self.assertAlmostEqual(f.interpolate(1.0000000001), 100, 6)
# Function with double-point (infinite derivative)
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
f.append(1, 0)
f.append(1, 100)
f.append(2, 100)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(1), 100, 6)
self.assertAlmostEqual(f.interpolate(-1), 0, 6)
# A more complex one
f = ContinousPiecewiseLinearFunc()
f.append(10, 9000)
f.append(20, 10000)
f.append(1000, 10980)
self.assertAlmostEqual(f.interpolate(5), 9000.0, 6)
self.assertAlmostEqual(f.interpolate(15), 9500.0, 6)
self.assertAlmostEqual(f.interpolate(40), 10020.0, 3)
self.assertAlmostEqual(f.interpolate(1010), 10980.0, 3)
def test_piecewise_linear_interpolation(self):
# Empty function
f = ContinousPiecewiseLinearFunc()
catched = False
try:
self.assertAlmostEqual(f.interpolate(0), 0, 3)
except:
catched = True
self.assertTrue(catched)
# Single point function
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(-1), 0, 6)
self.assertAlmostEqual(f.interpolate(1), 0, 6)
# A simple well-behaved function
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
f.append(1, 100)
self.assertAlmostEqual(f.interpolate(-0.000000001), 0, 6)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(0.000000001), 0, 6)
self.assertAlmostEqual(f.interpolate(0.5), 50.0, 6)
self.assertAlmostEqual(f.interpolate(0.9999999999), 100, 6)
self.assertAlmostEqual(f.interpolate(1.0), 100, 6)
self.assertAlmostEqual(f.interpolate(1.0000000001), 100, 6)
# Function with double-point (infinite derivative)
f = ContinousPiecewiseLinearFunc()
f.append(0, 0)
f.append(1, 0)
f.append(1, 100)
f.append(2, 100)
self.assertAlmostEqual(f.interpolate(0), 0, 6)
self.assertAlmostEqual(f.interpolate(1), 100, 6)
self.assertAlmostEqual(f.interpolate(-1), 0, 6)
# A more complex one
f = ContinousPiecewiseLinearFunc()
f.append(10, 9000)
f.append(20, 10000)
f.append(1000, 10980)
self.assertAlmostEqual(f.interpolate(5), 9000.0, 6)
self.assertAlmostEqual(f.interpolate(15), 9500.0, 6)
self.assertAlmostEqual(f.interpolate(40), 10020.0, 3)
self.assertAlmostEqual(f.interpolate(1010), 10980.0, 3)