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

When navigation guides drivers onto blocked roads, the inability to detect road closures hampers timely arrival; monitoring sudden drops in vehicle flow can reveal such events, but pedestrian and bicycle traffic introduce noise that obscures the signal.

To filter this noise, the study treats user navigation mode as a pseudo‑label for trajectory classification, acknowledging severe class imbalance and label inaccuracy, and proposes a label‑probability mixture Bayesian model to mitigate both issues.

01 Sample Acquisition and Label Definition

Because raw trajectories lack ground truth, the navigation mode (e.g., car navigation) is used as a proxy label; however, the distribution is heavily skewed toward cars and may not reflect actual travel mode.

02 Feature Analysis

Five groups of features are extracted: trajectory overview (duration, length, start time), speed‑related (max, mean, std), time‑related (red‑light wait, U‑turn, left‑turn times), behavior‑related (U‑turn, repeated activity, left‑turn vs right‑turn time), and user profile (occupation, car‑ownership probability).

Example: the probability density of start‑time shows higher car probability in the morning peak (5:00‑10:00), higher bicycle probability at noon (11:00‑13:00) due to food‑delivery, and higher pedestrian/bicycle probability in the evening (17:00‑20:00).

03 Bayesian Model from a Probability‑Distribution Perspective

The Bayesian classifier is chosen because it is a generative model with clear statistical meaning, allowing analysis of each feature’s probability density.

Key advantages:

Explicit conditional probability density functions enable feature effectiveness analysis.

Model can be expressed as a simple probabilistic formula (shown in the figure).

The label‑probability mixture Bayesian model iteratively refines pseudo‑labels using a maximum‑likelihood loss, cleaning noisy samples before re‑training the classifier.

Joint probability density functions for distance and max speed are constructed to replace the independence assumption, reducing misclassification of long‑distance bicycle or delivery‑person trajectories as cars.

04 Deep‑Learning Model from an Image‑Encoding Perspective

Trajectory data are encoded into 256×256 images where speed, direction, and waiting time are mapped to HSV channels (V = speed, H = direction, S = waiting time), preserving spatial distribution.

A dual‑stream network combines convolutional features from the encoded images with a static 1‑D feature vector (road‑type ratios, scaling factor) and feeds them into a fully‑connected layer; ResNet‑50 is used as the backbone.

05 Experimental Results

Both models were evaluated on ~100 manually labeled samples. The Bayesian mixture model achieved higher classification accuracy than the deep‑learning approach, likely because ResNet‑50 is not optimal for trajectory images, the training set lacked diversity, and the sampled scenarios differed from typical deep‑learning datasets.

Conclusion

Accurate trajectory classification is crucial for timely detection of road closures. The paper presents two technical solutions—probability‑based Bayesian modeling and image‑based deep learning—each with strengths and limitations, and suggests future improvements via scenario‑focused optimization and richer upstream data.

References

Frénay B, Verleysen M. Classification in the presence of label noise: a survey. IEEE Trans. Neural Netw. Learn. Syst., 2013.

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

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