2024/ml_accident_detection.py
●
ML Accident Detection System
Frame-level classification pipeline distinguishing “Accident” from “Non-Accident” scenarios reaching ~82% validation accuracy on real-time highway surveillance streams.
Python
OpenCV
TensorFlow
Scikit-Learn
01 // Input
CCTV Video Feed
RTSP (1080p@25)
Latency: ~2.1ms
▶
02 // Buffer
Frame Ring-IO
[1, 224, 224, 3]
OpenCV Normalization
▶
03 // Core
CNN Classifier
Softmax Logits [1, 2]
Inference: ~14.2ms
▶
04 // Filter
Temporal Window
P(Accident) > 0.70
Rolling Buffer (k=5)
▶
05 // Action
Incident Alert
MQTT / Webhook
Val Acc: ~82%
Target Deployment
Edge Jetson / POSIX Linux Server / RTSP camera hub
Memory Ceiling
< 180 MB RSS RAM footprint during 1080p ingestion
Concurrency Model
Threaded ring buffer decouple capture from inference
Fault Recovery
Automatic RTSP reconnect & frame-skip on jitter spike
Validation Accuracy
~82.4% on highway test video corpus
Throughput
25+ FPS real-time processing capability
False Positive Rate
< 4.1% via temporal sliding window confirmation
Pipeline Latency
End-to-end incident trigger < 85ms
Source Repository ↗
v1.4.2 [STABLE]