Skip to content
Back to search results

GPS Trajectory Analysis

Introduced by
Yu Zheng
Year
2015
Type
Pipeline for turning raw movement traces into structured mobility information
Family
Movement and trajectory data mining
Font size

What does it do?

GPS trajectory analysis is the pipeline that turns raw streams of timestamped location fixes into structured, meaningful mobility information — the stops where a person dwells, the trips between them, the transport modes used, and the network routes actually travelled.

Following the trajectory-data-mining framework synthesized by Yu Zheng in 2015, it cleans noisy positions, segments movement into stays and journeys, snaps points onto road or transit networks, and infers behaviour and recurrent patterns. It is the foundation for activity-space, travel-demand, and mobility studies built on smartphone and vehicle tracking data.

The intuition behind the method

A phone or tracker spits out a jittery breadcrumb trail of latitude–longitude points, some of them plainly wrong because a building bounced the signal. Looked at raw, it is just noise; but step back and structure emerges — long clusters of points where someone sat still (home, the office, a café) separated by stretches where the points march steadily along a road. The job of trajectory analysis is to read that trail the way a person would: throw out the impossible jumps, recognize the sitting-still places as stops, recognize the marches as trips, lay each trip onto the actual streets it must have followed, and judge from the speed whether the traveller was walking, cycling, or driving.

History

Trajectory analysis grew with the spread of GPS-enabled devices in the 2000s, drawing on moving-object database research, spatial data mining, and the hidden Markov map-matching methods formalized by Newson and Krumm in 2009. Yu Zheng's projects at Microsoft Research Asia — including the GeoLife GPS dataset and work on stay-point detection, mode inference, and route recommendation — were highly influential, and his 2015 ACM TIST survey 'Trajectory data mining: an overview' consolidated the field into the now-standard preprocessing, segmentation, matching, and inference pipeline. The subsequent ubiquity of smartphones and ride-hailing and fleet telematics turned trajectory mining into a core method across transport, urban computing, and mobility-aware human geography.

When to use it

Use GPS trajectory analysis whenever the raw material is sequences of location fixes from phones, vehicles, wearables, or fleets and you need to recover where people stopped, how they travelled, and along which routes. It is essential for building activity spaces, estimating travel demand and mode share, mapping congestion, or studying mobility behaviour from passively collected positioning data. It is not the right tool when movement is already known at the activity level (e.g. clean travel diaries), when positioning data are too sparse or coarse to recover stops and routes, or when only aggregate origin–destination flows are needed and individual traces add no value.

Limitations and considerations

  • Output quality is bounded by positioning accuracy, sampling rate, and gaps; sparse or urban-canyon data degrade every stage.
  • Stop detection and mode inference depend on thresholds and features that often need local calibration.
  • Map matching can fail on dense, multi-level, or poorly mapped networks, mis-assigning routes.
  • Continuous tracking raises serious privacy concerns, constraining collection, sharing, and reuse of trajectories.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Extracts rich, behaviour-level information — stops, trips, modes, routes — from cheap, passively collected location data.
  • Modular pipeline whose stages can be tuned or swapped independently for different data and questions.
  • Map matching grounds noisy points in the real network, enabling accurate distance, route, and exposure estimates.
  • Scales from individual diaries to population-scale mobility mining with mature open-source tooling.

Applications

  • Detecting home, work, and other significant locations and durations from smartphone traces.
  • Estimating travel mode share and trip patterns for transport planning from passive GPS data.
  • Map-matching fleet or ride-hailing traces to measure travel times, routes, and congestion.
  • Building individual activity spaces and mobility exposures for health and social research.
  • Mining recurrent routes and anomalies for navigation, logistics, and behaviour analytics.

How it works

How it works

1

Raw fixes are noisy: positions drift, signals are lost, and sampling rates vary. Preprocessing cleans the trace by removing implausible points — those implying impossible speeds or carrying poor accuracy flags — smoothing or filtering jitter, and optionally resampling to a regular interval. The aim is a trajectory faithful enough that downstream segmentation and matching are not corrupted by outliers, while preserving genuine movement.

2

The cleaned trajectory is split into stops and moves by detecting stay points: a stay point is a run of fixes that stay within a spatial radius r for at least a minimum duration τ, indicating the object lingered rather than passed through. The runs that satisfy this become stops (activity locations), and the segments connecting them become moves (trips). Tuning r and τ trades off catching brief stops against fragmenting a single visit.

Formula
\text{stay point} \iff \operatorname{dist}(p_i, p_j) \le r \;\text{and}\; (t_j - t_i) \ge \tau
3

Because GPS points rarely fall exactly on a road, map matching snaps the trajectory onto the underlying network. The widely used hidden Markov approach treats true road segments as hidden states and observed points as noisy emissions, then finds the road sequence that maximizes the joint likelihood of the observations and of plausible segment-to-segment transitions. The result is a route along real streets or rails rather than a floating line of points.

Formula
\hat{R} = \arg\max_{R} \prod_{k} P(o_k \mid r_k)\,P(r_k \mid r_{k-1})
4

Finally, behaviour is inferred from the structured trajectory. Transport mode is classified from kinematic features — speed, acceleration, stop frequency — that distinguish walking, cycling, bus, and car. Across many trajectories, pattern-mining methods extract recurrent behaviour such as frequent routes, periodic visits, and shared movement, turning individual traces into population-level mobility knowledge.

Formula
\hat{m} = \arg\max_{m} P(m \mid v, a, \dots)

Common pitfalls

  • Skipping preprocessing, so a single spurious jump fix creates a phantom high-speed trip or a false stop.
  • Setting the stay-point radius and duration without calibration, either merging distinct visits or splitting one stay into many.
  • Trusting map-matched routes on sparse traces, where the most likely path may diverge from the one actually taken.
  • Inferring mode from speed alone, confusing a slow-moving car in traffic with a cyclist or a stationary bus with a stop.

Frequently asked questions

What is a stay point and how is it detected?

A stay point is a location where the moving object lingered rather than passed through — an inferred stop such as home, work, or a shop. It is detected by scanning the trajectory for a run of consecutive fixes that remain within a spatial radius r of one another for at least a minimum duration τ. The run collapses to a single stay point at its centroid, and the segments between stay points become trips.

Why is map matching necessary?

GPS positions carry error and almost never fall exactly on a road, so a raw trajectory is a line of off-network points that cannot be tied to specific streets. Map matching snaps the trace onto the network — most robustly with a hidden Markov model that balances how close each point is to candidate segments against how plausible the transitions between segments are — yielding the route actually travelled, which is required for accurate distances, travel times, and route analysis.

How is transport mode inferred from a trajectory?

Mode is inferred from kinematic features computed along the trajectory — average and peak speed, acceleration, the frequency and length of stops, and sometimes proximity to transit lines. A classifier learns to map these features to modes such as walking, cycling, bus, or car, since each has a characteristic signature. Accuracy improves when features are combined and when network and timetable context disambiguate, for example, a bus from a car in similar traffic.

Related methods

Sources

  1. Zheng, Y. (2015). Trajectory data mining: an overview. ACM Transactions on Intelligent Systems and Technology, 6(3), 1–41.

    Open DOI reference
    Copied.