diff --git a/scripts/geodata/osm/intersections.py b/scripts/geodata/osm/intersections.py
new file mode 100644
index 00000000..831539da
--- /dev/null
+++ b/scripts/geodata/osm/intersections.py
@@ -0,0 +1,102 @@
+import array
+import logging
+import six
+
+from bisect import bisect_left
+from collections import defaultdict, OrderedDict
+from itertools import izip, combinations
+
+from geodata.coordinates.conversion import latlon_to_decimal
+from geodata.osm.extract import *
+
+
+class OSMIntersectionReader(object):
+    def __init__(self, filename):
+        self.filename = filename
+
+        self.node_ids = array.array('l')
+        self.node_coordinates = array.array('d')
+
+        # Store these in memory, could be LevelDB if needed
+        self.way_props = {}
+        self.intersections_graph = defaultdict(list)
+
+
+    def binary_search(self, a, x):
+        '''Locate the leftmost value exactly equal to x'''
+        i = bisect_left(a, x)
+        if i != len(a) and a[i] == x:
+            return i
+        return None
+
+    def intersections(self):
+        '''
+        Generator which yields tuples like:
+
+        (node_id, lat, lon, {way_id: way_props})
+        '''
+        i = 0
+
+        node_ids = array.array('l')
+        node_counts = array.array('i')
+
+        for element_id, props, deps in parse_osm(self.filename, dependencies=True):
+            props = {safe_decode(k): safe_decode(v) for k, v in six.iteritems(props)}
+            if element_id.startswith('node'):
+                node_id = long(element_id.split(':')[-1])
+                node_ids.append(node_id)
+                node_counts.append(0)
+            elif element_id.startswith('way'):
+                # Don't care about the ordering of the nodes, and want uniques e.g. for circular roads
+                deps = set(deps)
+
+                # Get node indices by binary search
+                try:
+                    node_indices = [self.binary_search(node_ids, node_id) for node_id in deps]
+                except ValueError:
+                    continue
+
+                # way_deps is the list of dependent node ids
+                # way_coords is a copy of coords indexed by way ids
+                for node_id, node_index in izip(deps, node_indices):
+                    node_counts[node_index] += 1
+
+            if i % 1000 == 0 and i > 0:
+                print('doing {}s, at {}'.format(element_id.split(':')[0], i))
+            i += 1
+
+        for i, count in enumerate(node_counts):
+            if count > 1:
+                self.node_ids.append(node_ids[i])
+
+        del node_ids
+        del node_counts
+
+        i = 0
+
+        for element_id, props, deps in parse_osm(self.filename, dependencies=True):
+            if element_id.startswith('node'):
+                node_index = self.binary_search(self.node_ids, node_id)
+                if node_index is not None:
+                    lat = props.get('lat')
+                    lon = props.get('lon')
+                    lat, lon = latlon_to_decimal(lat, lon)
+                    self.node_coordinates.extend([lat, lon])
+            elif element_id.startswith('way'):
+                props = {safe_decode(k): safe_decode(v) for k, v in six.iteritems(props)}
+                way_id = long(element_id.split(':')[-1])
+                props['id'] = way_id
+                for node_id in deps:
+                    node_index = self.binary_search(self.node_ids, node_id)
+                    if node_index is not None:
+                        self.intersections_graph[node_index].append(way_id)
+                        self.way_props[way_id] = props
+
+            if i % 1000 == 0 and i > 0:
+                print('second pass, doing {}s, at {}'.format(element_id.split(':')[0], i))
+            i += 1
+
+        for node_index, way_indices in six.iteritems(self.intersections_graph):
+            lat, lon = self.node_coordinates[node_index * 2], self.node_coordinates[node_index * 2 + 1]
+            ways = [self.way_props[w] for w in way_indices]
+            yield self.node_ids[node_index], lat, lon, ways