Counting cars is not measuring traffic.

A congestion sensor built entirely out of a CCTV feed. Two fine-tuned YOLO11n models — one detecting vehicles, one segmenting the road surface — turn every frame into six numbers describing how traffic is behaving, and a small classifier reads those numbers as congested or not. 97.5% accuracy on the unseen test set, with the whole pipeline running at 18–22 FPS. My undergraduate thesis at Petra Christian University, defended in 2025 against live feeds from Malang City.

Malang is a hard case on purpose. Indonesian traffic is heterogeneous and motorcycle-dominated — motorcycles are over 83% of vehicles nationally — so a detector trained on car-shaped Western traffic has very little to say about it, and a model that can't see motorbikes can't see the jam.

Sole engineer. I collected and hand-annotated all three datasets, fine-tuned both YOLO11 models, wrote the feature-extraction layer that turns boxes and masks into traffic numbers, trained and compared the two classifiers, ran the evaluation — and defended the whole thing.

Congestion in Indonesian cities is measured badly or not at all, and it is expensive either way.

• The cost is national-scale. Congestion is estimated to cost IDR 63.4 trillion a year in lost economic activity and transport efficiency (DirJen Kemenhub, 2024).
• The instruments are the obstacle. Existing monitoring depends on costly, invasive hardware — in-ground loops, roadside sensors — or trades away accuracy to run in real time (Cui et al., 2020). Most intersections therefore go unmeasured.
• Vision systems take the easy signal. Earlier camera-based work classifies on vehicle count alone, ignoring density, occupancy and speed. Twenty vehicles moving freely and twenty vehicles stopped are the same count.

So the thesis had two questions to answer: does classifying on four traffic-flow features — flow, occupancy, density and speed — detect congestion more accurately than counting? And how accurate is YOLO11 in the first place on a vehicle population that is mostly motorbikes?

Measure behaviour, not headcount. The detector's job is not to answer the question — it is to produce the raw material the four features are computed from.

The four features

Road segmentation is what makes density and occupancy mean anything. Without it, density is vehicles-per-frame, which changes with camera angle rather than with traffic.

Three datasets, all hand-labelled

Two branches of a single frame meet again at the feature vector: the detector says what and how many, the segmenter says out of how much road, and the motion between frames says how fast. Only then does anything classify.

How each model was trained

The same shape ran three times — YOLO11n for detection, YOLO11n-seg for segmentation, and the classifiers over the extracted features. Only the dataset and the metrics change; the SVM adds a grid search over its hyperparameters in the training step.

Speed is measured using Lucas-Kanade sparse optical flow — tracking the pixel displacement of detected vehicle centroids frame-to-frame and converting it to a normalised speed feature. Combined with SSIM-based frame sampling to skip visually redundant frames, this keeps the pipeline at 18–22 FPS on real CCTV footage.

Vehicle detection — YOLO11n

Cars are close to solved at 95.2% mAP@50. Motorbikes are the honest number: 75.4% recall, and mAP@50-95 of 56.4%. They are small, they cluster, and in a queue they physically occlude one another — which is exactly the condition the system is meant to detect.

Road segmentation — YOLO11n-seg

Decent in daylight and clearly the weaker half of the pipeline at night: glare and uneven street lighting break the road mask into fragments, and everything computed against road area degrades with it.

Feature extraction

The raw YOLO outputs — boxes and masks — are converted per frame into the six numerical features, printed alongside the overlay while the stream runs. This is the layer that made the dataset: 20,113 rows of real traffic state.

The classifier — two of them

Both were trained on the same 20,113 rows and both land within two-tenths of a percent of each other, which says more about the features than about either model. The neural network is the one the conclusion runs on.

TrafficClassifier(
  (layer_1):     Linear(in_features=6, out_features=64, bias=True)
  (relu1):       ReLU()
  (dropout1):    Dropout(p=0.3, inplace=False)
  (layer_2):     Linear(in_features=64, out_features=32, bias=True)
  (relu2):       ReLU()
  (dropout2):    Dropout(p=0.3, inplace=False)
  (output_layer):Linear(in_features=32, out_features=1, bias=True)
  (sigmoid):     Sigmoid()
)

Six inputs, two hidden layers of 64 and 32 units, ReLU with 0.3 dropout on each, one sigmoid output. Small on purpose — the features carry the signal, so the model doesn't have to.

Two numbers in the results are worse than the headline, and both were in the defence rather than hidden behind it.

And a scope limit worth stating plainly: every number here is measured on Malang City CCTV. The thesis makes no claim about other cities, other camera heights, or other traffic mixes.

• Python
• YOLO11
• PyTorch
• OpenCV
• Lucas-Kanade flow
• scikit-learn · SVM
• SSIM sampling
• Roboflow

The feature engineering beat the bigger model. Counting vehicles is the obvious signal and the wrong one — twenty moving and twenty stopped are the same count. Once flow, density, occupancy and speed were right, two hidden layers were enough to separate them, and a grid-searched SVM landed within two-tenths of a percent of the network. The work was never in the classifier.