When we view a vehicle's route or a pedestrian's path on a digital map, we expect a clean, continuous line that perfectly traces the underlying infrastructure. In reality, raw spatial data collected from GPS sensors, cellular networks, or radar is fundamentally messy. It jitters, drifts, skips updates, and occasionally implies that a vehicle is driving through a building or flying across a river.
Because downstream applications—such as route prediction, travel time estimation, and urban planning—rely on precise spatial inputs, handling this inherent noise is a foundational challenge in trajectory data management. Modern frameworks solve this not by simply erasing bad data, but by leveraging physical constraints, sequential probabilities, and structural context to reconstruct the true path.
The Physics of Trajectory Noise
To clean trajectory data effectively, we must first understand why it breaks. In urban environments, the primary culprit behind GPS degradation is the "urban canyon" effect. Tall buildings block direct line-of-sight signals from satellites, causing radio waves to bounce off concrete and glass before reaching a receiver. This phenomenon, known as multipath interference, introduces significant positional offsets and artificial jumps (Prabowo et al.; Wang).
Furthermore, hardware-level smoothing can introduce its own artifacts. For instance, smartphone internal Kalman filters can over-correct during sharp turns, creating artificial loops or straight-line overshoots (Apple et al.). When tracking relies on cellular network data (such as GSM cell IDs) rather than GPS, the noise manifests as "cell oscillation" or "station drift," where a stationary device rapidly jumps between nearby towers, creating a false illusion of high-speed movement (Shad et al.; Wang et al.).
Why Simple Proximity Fails
The most intuitive approach to cleaning a trajectory is geometric proximity: taking each noisy coordinate and "snapping" it to the nearest road segment. However, this naive strategy fails in complex urban layouts. If a GPS coordinate drifts by 20 meters, it may land closer to a parallel side street or an elevated overpass than the highway the vehicle is actually traversing (Imamichi et al.; Sun).
Treating each data point as an isolated coordinate ignores the temporal continuity of motion. Snapping points independently results in a reconstructed path that zig-zags erratically between disconnected streets, implying physically impossible maneuvers and rendering the data useless for traffic analysis (Botteldooren et al.).
Enforcing Kinematic Constraints
To reconstruct a plausible path, algorithms must enforce the laws of physics. Vehicles possess mass, maximum acceleration limits, and strict directional constraints—they cannot instantly teleport or move backward on a freeway.
One way to eliminate high-frequency jitter is through kinematic smoothing. For example, the authors of a study on public transit tracking combined local regression with monotonic cubic splines to process high-frequency GPS data (Huang et al.). By forcing the underlying mathematical curve to be monotonic, the model ensures that the reconstructed vehicle always moves forward, smoothing out impossible velocity spikes caused by sensor drift.
Similarly, hybrid pipelines can combine limiting filters (to clip extreme outliers) with moving averages and downsampling algorithms like Douglas–Peucker (Wang). This dual action strips away high-frequency noise while compressing the dataset into its essential structural vertices.
Raw GPS Pings:
* * * (High jitter, off-road drift)
\ / /
Kinematic Filter: ------*----- (Enforces smooth velocity & splines)
|
Map Topology: ================ (Snaps to physically reachable road)
Sequential State Estimation
Instead of evaluating points individually, robust map-matching frameworks model the entire trajectory as a continuous sequence using Hidden Markov Models (HMMs). This approach balances two competing probabilities:
- Emission Probability: How close is the noisy GPS point to a candidate road segment?
- Transition Probability: How physically plausible is the route between the previous candidate segment and the current candidate segment?
If a GPS point drifts onto a parallel alleyway, the emission probability might be high, but the transition probability will be near zero if there is a concrete wall separating the highway from that alley. The HMM looks at the path holistically, selecting the sequence of road segments that maximizes the overall joint probability (Imamichi et al.; Sun).
When data is highly sparse—such as commercial fleet tracking where pings occur minutes apart—intermediate points are entirely missing. To resolve this, the focus shifts from point-to-line matching to "spatial semantic features" (Luo et al.). By evaluating macro-level attributes like the number of turns and changes in road categories across the entire sequence, the true path can be inferred even when the underlying coordinate data is incomplete.
Multimodal Representation Learning
Modern deep learning architectures move beyond hand-crafted geometric rules by learning latent representations of trajectories that are inherently robust to noise. The core insight is that a trajectory can be viewed as a multimodal observation consisting of high-resolution micro-motion (raw coordinates, instantaneous speed, heading) and macro-level semantics (the underlying road network topology or spatial grid expressions).
The GREEN framework, for instance, utilizes a dual-encoder architecture to fuse grid expressions and road network paths (Shang et al.). If environmental interference causes a local signal inversion or a sequence of missing pings, the grid encoder maintains the broader spatial context while the road encoder enforces topological continuity.
Similarly, the JGRM framework jointly models raw GPS traces and map-matched routes (Ma et al.). By anchoring noisy micro-motions to stable macro-semantic structures during training, the model learns low-dimensional embeddings that capture the true underlying travel behavior. Consequently, downstream tasks like travel time estimation and similarity searches become highly accurate, as the neural network learns to ignore local sensor noise in favor of global structural patterns.
Key Foundational Work
From GPS traces to a routable road map (2009) — contributed 2 concept elements to this direction.
Route Prediction from Trip Observations (2008) — contributed 2 concept elements to this direction.