Speed Estimation from Road Scene Sequences

H. Choi

Introduction to Artificial Intelligence Course Project

This project focuses on estimating vehicle speed from sequential road scene images using a combination of 3D-CNN and LSTM architectures. The model processes 6-frame sequences to predict speed with high accuracy, incorporating various optimization techniques including dropout, learning rate schedulers, and test-time normalization.

Project Page

Model Architecture

3D-CNN: Spatial-temporal feature extraction from 6-frame sequences
LSTM: Temporal sequence modeling
Input: 6 consecutive road scene frames
Output: Speed estimation (regression)

Optimization Techniques

Dropout regularization
Learning rate scheduling
Test-time normalization
Data augmentation

Performance

Mean Absolute Error (MAE): 1.82
Effective spatiotemporal feature learning
Robust performance across different road conditions