Trajectory Classification for Road Closure Detection Using Bayesian and Deep Learning Approaches

1. Background

When using navigation, drivers may encounter situations where the map suggests a route that is actually blocked due to construction, leading to navigation failures. Detecting dynamically which roads are impassable is crucial for accurate routing.

Road closures cause a sudden drop in vehicle traffic on the affected road. However, pedestrian and bicycle traffic can mask this change, making it difficult to identify closures based solely on overall traffic flow.

Classifying trajectories of vehicles, bicycles, and pedestrians allows us to filter out noise from non‑vehicle traffic and focus on vehicle flow, making road‑closure signals more pronounced.

2. Sample Acquisition and Label Definition

Since raw trajectory data lack ground‑truth labels, the user’s navigation mode is used as a pseudo‑label: trajectories generated while the user follows car navigation are labeled as "car", etc. This creates a severe class‑imbalance (car data far outnumber non‑motorized data) and introduces label noise because navigation mode may not match the actual travel mode. A probability‑mixture Bayesian model is proposed to address both issues.

3. Feature Analysis

Five groups of features are extracted:

Trajectory overview features (duration, length, start time, etc.).

Speed‑related features (max speed, average speed, speed standard deviation).

Time‑related features (waiting at red lights, U‑turn time, left‑turn time).

Behavior‑related features (U‑turn behavior, repeated activity, left‑turn vs right‑turn time).

User profile features (occupation, car‑ownership probability).

Example: the start‑time feature shows distinct probability density functions for car, bicycle, and walking trajectories at different times of day.

4. Bayesian Model Perspective

Reasons for choosing a Bayesian classifier:

It is a generative model with clear statistical meaning, allowing analysis of each feature via conditional probability density functions.

The classifier can be expressed as P(C|X) = \frac{P(X|C)P(C)}{P(X)} , enabling easy iteration by adding or removing features.

Its output is a probability (confidence) rather than a hard label.

4.1 Label‑Probability Mixture Bayesian Model

The problem involves data imbalance and noisy labels. Inspired by Tanaka et al.’s pseudo‑label loss, a pseudo‑label maximum likelihood estimation loss is incorporated into the Bayesian framework to clean noisy samples before iterating the classifier.

4.2 Joint Probability Density Function Calculation

Assuming independence between features can cause misclassifications (e.g., long‑distance, moderate‑speed bicycle trips being labeled as cars). By constructing joint probability density functions for distance and speed, the model better distinguishes vehicle types.

4.3 Experimental Results of Bayesian Classification

Approximately 100 sampled trajectories were manually annotated. The Bayesian model achieved higher classification accuracy compared to baseline methods (details shown in the accompanying tables).

5. Deep Learning Model from Image Encoding Perspective

Deep learning can learn directly from raw data without manual feature engineering. Two ways to observe trajectories are considered:

Time series : GPS points uploaded sequentially; models such as TCN or RNN can capture temporal patterns.

Spatial distribution : Trajectories plotted on a map as images; encoding speed, direction, and waiting time into color channels enables CNNs to learn spatial patterns.

Speed, direction, and waiting time are encoded using the HSV color space (V = speed, H = direction, S = waiting time) and resized to 256×256 images.

5.2 Dual‑Stream Neural Network

In addition to the encoded trajectory image, a 1‑D static feature vector (e.g., proportion of road types traversed, scaling factor) is concatenated with CNN‑extracted features. The combined representation is fed into a fully connected layer for classification. ResNet‑50 is used as the backbone CNN.

5.3 Deep Learning Experimental Results

Using about 100 manually labeled samples, the deep learning model achieved reasonable accuracy but was still inferior to the Bayesian approach, likely due to limited training data, suboptimal model choice (ResNet‑50 may not be ideal for trajectory images), and distribution mismatch between training and real‑world scenarios.

6. Conclusion

Trajectory classification is vital for timely detection of road closures. This article explored both Bayesian models with probability density analysis and deep‑learning models with image encoding. Future work includes tailoring models to specific application scenarios, enriching upstream data, and improving feature engineering.

References

[1] Frénay B, Verleysen M. Classification in the presence of label noise: a survey. IEEE Transactions on Neural Networks and Learning Systems, 2013.

[2] Tanaka D, et al. Joint optimization framework for learning with noisy labels. CVPR, 2018.

[3] Liu Y, Wang X, You W. Non‑intrusive Load Monitoring by Voltage‑Current Trajectory Enabled Transfer Learning. IEEE Transactions on Smart Grid, 2018.