gtfs-segmenter 0.1.0

Segment GTFS shapes into stop-to-stop route segments with loop-aware projection

gtfs-segments

Segment GTFS shapes into stop-to-stop route segments that follow the original road geometry.

Most GTFS tools use Shapely's .project() to snap stops to shapes, which breaks on routes with loops, lollipop turnarounds, and out-and-back corridors. This library uses sequential projection with candidate-aware selection to handle complex route geometries correctly.

Install

pip install gtfs-segments

For interactive map visualization:

pip install gtfs-segments[viz]

Quick Start

Segment an entire feed

from gtfs_segments import segment_feed

result = segment_feed("path/to/gtfs")

segments = result['segments']       # GeoDataFrame of all stop-to-stop segments
diagnostics = result['diagnostics'] # per-shape stats (offsets, degenerates)
failures = result['failures']       # any shapes that failed

Segment a single shape

import pandas as pd
from gtfs_segments import segment_shape

shapes = pd.read_csv("gtfs/shapes.txt")
trips = pd.read_csv("gtfs/trips.txt")
stops = pd.read_csv("gtfs/stops.txt")
stop_times = pd.read_csv("gtfs/stop_times.txt")

result = segment_shape(shapes, trips, stops, stop_times, shape_id=232864)

segments = result['segments']
projected = result['projected_stops']
ruler = result['ruler']
diag = result['diagnostics']

Low-level control

from gtfs_segments import ShapeRuler, project_stops_sequential, segment_route

ruler = ShapeRuler(shapes, shape_id=232864)

# Get stop visit order from stop_times
trip_id = trips[trips['shape_id'] == 232864]['trip_id'].iloc[0]
st = stop_times[stop_times['trip_id'] == trip_id].sort_values('stop_sequence')
stop_sequence = st['stop_id'].tolist()
route_stops = stops[stops['stop_id'].isin(stop_sequence)].copy()

# Project stops onto shape and build segments
projected = project_stops_sequential(ruler, route_stops, stop_sequence)
segments = segment_route(ruler, projected)

How It Works

The Problem

Shapely's .project() finds the globally nearest point on a LineString. On a route that doubles back on itself, a stop at the base of a loop has two valid projection points — the entry and the exit. .project() picks whichever is geometrically closer, often the exit, which causes every stop inside the loop to collapse to the same distance.

The Solution

1. ShapeRuler converts the shape to a metric CRS, builds a cumulative distance ruler, and indexes every segment for fast searching.

2. find_candidates scans every segment and collects all positions where a stop projects within tolerance (default 50m). A stop at a loop base gets candidates at both entry and exit.

3. project_stops_sequential processes stops in visit order (from stop_times), enforcing monotonic distances. For each stop, it picks the earliest candidate with a reasonable offset — biasing toward forward progress along the route rather than jumping ahead to a later occurrence.

4. segment_route slices the shape between consecutive stop distances, including all original vertices so segments follow the actual road geometry.

CRS Auto-Detection

If no CRS is provided, the library auto-detects the UTM zone from the median longitude of the shape coordinates. Override with metric_crs="EPSG:32614" if needed.

Diagnostics

The diagnostics DataFrame flags potential issues:

• max_offset_m: Largest perpendicular distance between a stop and the shape. Over 30m usually means the stop is set back from the road (transit centers, park-and-rides) or assigned to the wrong shape.
• n_high_offset: Count of stops with offset > 30m.
• n_degenerate: Count of segments shorter than 5m. Usually near-side/far-side stop pairs at intersections.

Requirements

• Python ≥ 3.9
• geopandas ≥ 0.12
• shapely ≥ 2.0
• pandas ≥ 1.5
• numpy ≥ 1.23

License

MIT