All projects
Project 05 · ML · Deep Learning
Accident Detection Using CNNs
A convolutional neural network system that detects road accidents from both static images and live video streams in real time. The pipeline uses TensorFlow and Keras for model training, PyTorch for experimentation, and OpenCV for video frame extraction and inference — with an automated alert generation system that triggers immediately on accident detection to accelerate emergency response.
CNN model architecture
A multi-layer convolutional neural network trained to classify accident vs. non-accident frames from road imagery. Built and trained in TensorFlow/Keras with careful data augmentation and class imbalance handling — since accident footage is naturally rare in real-world datasets.
Real-time video inference
OpenCV handles video capture and frame-by-frame extraction from both live camera feeds and pre-recorded clips. Each frame is preprocessed and passed through the CNN at inference time, enabling continuous real-time monitoring of road conditions.
Automated alert pipeline
When the CNN detects an accident frame above the confidence threshold, an alert is generated and dispatched automatically — designed to integrate with emergency response systems and reduce the time between accident occurrence and first responder notification.
Dual-framework ML pipeline
TensorFlow/Keras handles production model training and export, while PyTorch is used for experimentation and architecture comparison. This dual-framework setup allows iterating on model design without being locked into a single ecosystem.