PedSense-AI: YOLO26 Pedestrian Intent Predictor
PedSense-AI is a computer vision framework for predicting pedestrian crossing intent. This repository hosts the YOLO26 End-to-End Detector, optimized for real-time inference and high-FPS applications.
Architectural Approach: YOLO26
Unlike temporal models, this version of YOLO26 is fine-tuned to classify crossing vs. not-crossing intent directly within the detection head.
- Type: End-to-end single-stage detector.
- Input: Single RGB frames.
- Best For: Real-time edge deployment where low latency is critical.
Performance (10-Epoch Benchmark)
Trained on the JAAD (Joint Attention in Autonomous Driving) dataset, the model achieves the following validation metrics:
| Metric | Score | Definition |
|---|---|---|
| mAP@50 | 0.631 | Mean Average Precision at 0.5 Intersection over Union (IoU). |
| mAP@50-95 | 0.453 | Average mAP across IoU thresholds from 0.5 to 0.95. |
| Precision | 0.684 | Probability that a detection is a true pedestrian. |
| Recall | 0.600 | Probability that a true pedestrian is detected. |
Training Dynamics
The model showed rapid convergence, with Classification Loss dropping from 2.25 to 0.74 over 10 epochs. This indicates a strong ability to distinguish intent categories even without temporal data.
Usage
from ultralytics import YOLO
# Load from Hugging Face
model = YOLO("hf://JcProg/pedsense-yolo26-jaad-10e.pt")
# Predict intent (Class 0: Not-Crossing, Class 1: Crossing)
results = model.predict("street_scene.jpg")
results[0].show()
- Downloads last month
- 2