app.py

import gradio as gr
import cv2
import numpy as np
import tempfile
import os
from scripts.inference import GazePredictor
from utils.ear_utils import BlinkDetector

def smooth_values(history, current_value, window_size=5):
    if current_value is not None:
        history.append(current_value)
    if len(history) > window_size:
        history.pop(0)
    return np.mean(history, axis=0) if isinstance(current_value, np.ndarray) and history else current_value if current_value is not None else 0

MODEL_PATH = os.path.join("models", "gaze_estimation_model.pth")

def analyze_video(input_video):
    cap = cv2.VideoCapture(input_video)
    gaze_predictor = GazePredictor(MODEL_PATH)
    blink_detector = BlinkDetector()
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    temp_fd, temp_path = tempfile.mkstemp(suffix='.mp4')
    os.close(temp_fd)
    out = None

    GAZE_STABILITY_THRESHOLD = 0.5
    TIME_THRESHOLD = 15
    BLINK_RATE_THRESHOLD = 1
    EYE_CLOSURE_THRESHOLD = 10
    HEAD_STABILITY_THRESHOLD = 0.05

    gaze_history = []
    head_history = []
    ear_history = []
    stable_gaze_time = 0
    stable_head_time = 0
    eye_closed_time = 0
    blink_count = 0
    start_time = 0
    is_unconscious = False

    frame_count = 0
    fps = cap.get(cv2.CAP_PROP_FPS) or 20

    while True:
        ret, frame = cap.read()
        if not ret:
            break
        frame_count += 1
        if start_time == 0:
            start_time = frame_count / fps

        head_pose_gaze, gaze_h, gaze_v = gaze_predictor.predict_gaze(frame)
        current_gaze = np.array([gaze_h, gaze_v])
        smoothed_gaze = smooth_values(gaze_history, current_gaze)

        ear, left_eye, right_eye, head_pose, left_iris, right_iris = blink_detector.detect_blinks(frame)
        if ear is None:
            cv2.putText(frame, "No face detected", (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
            smoothed_head = smooth_values(head_history, None)
            smoothed_ear = smooth_values(ear_history, None)
        else:
            smoothed_head = smooth_values(head_history, head_pose)
            smoothed_ear = smooth_values(ear_history, ear)
            if smoothed_ear >= blink_detector.EAR_THRESHOLD:
                cv2.drawMarker(frame, left_iris, (0, 255, 0), markerType=cv2.MARKER_CROSS, markerSize=10, thickness=2)
                cv2.drawMarker(frame, right_iris, (0, 255, 0), markerType=cv2.MARKER_CROSS, markerSize=10, thickness=2)

        cv2.putText(frame, f"Gaze H: {smoothed_gaze[0]:.2f}", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
        cv2.putText(frame, f"Gaze V: {smoothed_gaze[1]:.2f}", (10, 90), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
        cv2.putText(frame, f"Head Pose: {smoothed_head:.2f}", (10, 120), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
        cv2.putText(frame, f"EAR: {smoothed_ear:.2f}", (10, 150), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)

        if len(gaze_history) > 1:
            gaze_diff = np.sqrt(np.sum((smoothed_gaze - gaze_history[-2])**2))
            if gaze_diff < GAZE_STABILITY_THRESHOLD:
                if stable_gaze_time == 0:
                    stable_gaze_time = frame_count / fps
            else:
                stable_gaze_time = 0

        if len(head_history) > 1 and head_pose is not None:
            head_diff = abs(smoothed_head - head_history[-2])
            if head_diff < HEAD_STABILITY_THRESHOLD:
                if stable_head_time == 0:
                    stable_head_time = frame_count / fps
            else:
                stable_head_time = 0

        if ear is not None and smoothed_ear < blink_detector.EAR_THRESHOLD:
            if eye_closed_time == 0:
                eye_closed_time = frame_count / fps
            elif (frame_count / fps) - eye_closed_time > EYE_CLOSURE_THRESHOLD:
                cv2.putText(frame, "Eyes Closed", (10, 210), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
        else:
            if eye_closed_time > 0 and (frame_count / fps) - eye_closed_time < 0.5:
                blink_count += 1
            eye_closed_time = 0

        elapsed_minutes = ((frame_count / fps) - start_time) / 60 if start_time > 0 else 0
        blink_rate = blink_count / elapsed_minutes if elapsed_minutes > 0 else 0
        cv2.putText(frame, f"Blink Rate: {blink_rate:.1f}/min", (10, 240), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)

        unconscious_conditions = [
            stable_gaze_time > 0 and (frame_count / fps) - stable_gaze_time > TIME_THRESHOLD,
            blink_rate < BLINK_RATE_THRESHOLD and elapsed_minutes > 1,
            eye_closed_time > 0 and (frame_count / fps) - eye_closed_time > EYE_CLOSURE_THRESHOLD,
            stable_head_time > 0 and (frame_count / fps) - stable_head_time > TIME_THRESHOLD
        ]
        if sum(unconscious_conditions) >= 2:
            cv2.putText(frame, "Unconscious Detected", (10, 270), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
            is_unconscious = True
        else:
            is_unconscious = False

        if out is None:
            h, w = frame.shape[:2]
            out = cv2.VideoWriter(temp_path, fourcc, fps, (w, h))
        out.write(frame)
    cap.release()
    if out:
        out.release()
    return temp_path

iface = gr.Interface(
    fn=analyze_video,
    inputs=gr.Video(),
    outputs=gr.Video(),
    title="Gaze Tracker",
    description="Upload a video to analyze gaze and drowsiness."
)

if __name__ == "__main__":
    iface.launch()