detector-classifier

Predicting traffic signal detector phase and function from hi-res event logs, to aid in configuring detectors for ATSPMs. Work in progress.

Kaggle Competition

This repository contains the code for the Traffic Signal Detector Classifier competition on Kaggle. The datasets can be downloaded from the competition page.

Model Architecture

The detector classifier uses a deep learning architecture to predict both the phase and function of traffic signal detectors based on their activation patterns and signal phase timing data.

Input

• Sequence length: 300 timesteps
• Features per timestep: 10
  • 8 channels for phase states
  • 1 channel for detector state
  • 1 channel for time delta between events

Architecture Layers

Input Layer (300 timesteps × 10 features)
↓
Masking Layer (for handling padded sequences)
↓
Bidirectional LSTM Layer 1

64 units per direction return sequences enabled total 128 features per timestep output
↓
Dropout Layer (0.2)
↓
Bidirectional LSTM Layer 2 32 units per direction total 64 features output
↓
Dropout Layer (0.2)
↓
Dense Layer 64 units ReLU activation
↓
Output Branches: ├→ Phase Output │ - Dense layer (8 units) │ - Softmax activation │ - Predicts 8 possible phases │ └→ Function Output

• Dense layer (3 units)
• Softmax activation
• Predicts 3 possible functions (Advance, Presence, Count)

Loss Function

Combined weighted loss: total_loss = α * phase_loss + β * function_loss

where:

• phase_loss: Categorical crossentropy for phase prediction
• function_loss: Categorical crossentropy for function prediction
• α, β: Configurable weight parameters (default: α = β = 0.5)