3D ResNet Car Crash Detector (TorchScript)

This repository hosts a compiled TorchScript (JIT) model for automated traffic accident and car crash detection from video clips.

Because this model is exported via TorchScript, it has zero external codebase dependenciesโ€”you do not need to clone the training repository or import any custom PyTorch Lightning classes.


Model Specifications

  • Architecture: 3D ResNet-18 (r3d_18)
  • Input Format: (Batch, Channels, Frames, Height, Width) -> (1, 3, 16, 112, 112)
  • Color Space: RGB normalized to [0.0, 1.0]
  • Output: Single float tensor representing Crash Probability [0.0 to 1.0] (Sigmoid pre-applied).

Quickstart Usage

1. Install Dependencies

pip install torch torchvision opencv-python numpy huggingface_hub

2. Python Inference Script

Paste and run the following script to load the model directly from Hugging Face and evaluate any .mp4 or .avi video file:

import cv2
import numpy as np
import torch
from huggingface_hub import hf_hub_download

# Configuration
HF_REPO_ID = "HaniaRuby/crash-detector-3dresnet"
HF_FILENAME = "crash_detector_3dresnet_jit.pt"

def preprocess_video(video_path, num_frames=16, frame_size=(112, 112)):
    """Extracts 16 uniform frames from a video file, normalizes, and reshapes for 3D ResNet."""
    cap = cv2.VideoCapture(video_path)
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    if total_frames <= 0:
        raise ValueError(f"Could not read frames from video file: {video_path}")

    # Sample 'num_frames' evenly across the clip duration
    frame_indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
    frames = []

    for idx in range(total_frames):
        ret, frame = cap.read()
        if not ret:
            break
        if idx in frame_indices:
            # Convert OpenCV BGR -> RGB, resize, and normalize to [0, 1]
            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            frame = cv2.resize(frame, frame_size)
            frame = frame.astype(np.float32) / 255.0
            frames.append(frame)

    cap.release()

    # Zero-pad if video duration is under 16 frames
    while len(frames) < num_frames:
        frames.append(np.zeros((*frame_size, 3), dtype=np.float32))

    # Convert to NumPy array: (16, 112, 112, 3)
    video_np = np.array(frames[:num_frames], dtype=np.float32)
    
    # Transpose to PyTorch 3D CNN tensor shape: (C, T, H, W) -> (3, 16, 112, 112)
    video_tensor = torch.tensor(video_np).permute(3, 0, 1, 2)
    
    # Add Batch dimension -> (1, 3, 16, 112, 112)
    return video_tensor.unsqueeze(0)

def predict_crash(video_path):
    print(f"Fetching compiled model from Hugging Face: {HF_REPO_ID}...")
    model_path = hf_hub_download(repo_id=HF_REPO_ID, filename=HF_FILENAME)

    print("Loading compiled TorchScript runtime...")
    model = torch.jit.load(model_path)
    model.eval()

    print(f"Processing video file: {video_path}")
    input_tensor = preprocess_video(video_path)

    print("Running model inference...")
    with torch.no_grad():
        crash_probability = model(input_tensor).item()

    print("\n" + "="*45)
    print(f"Input Video       : {video_path}")
    print(f"Crash Probability : {crash_probability * 100:.2f}%")
    print(f"Status            : {'CRASH DETECTED' if crash_probability > 0.5 else 'NORMAL DRIVING'}")
    print("="*45)

if __name__ == "__main__":
    # Replace with path to your video file
    video_file = "path/to/sample_video.mp4"
    predict_crash(video_file)
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support