import streamlit as st
import cv2
import torch
from PIL import Image
import numpy as np
from transformers import BlipProcessor, BlipForConditionalGeneration
from transformers import ViltProcessor, ViltForQuestionAnswering
import time
from io import BytesIO
import threading
import queue
import os
import tempfile
from datetime import datetime
# Set page config to wide mode
st.set_page_config(layout="wide", page_title="Securade.ai Sentinel")
def initialize_state():
if 'initialized' not in st.session_state:
st.session_state.frame = None
st.session_state.captions = []
st.session_state.stop_event = threading.Event()
st.session_state.frame_queue = queue.Queue(maxsize=1)
st.session_state.caption_queue = queue.Queue(maxsize=10)
st.session_state.processor = None
st.session_state.thread = None
st.session_state.is_streaming = False
st.session_state.initialized = True
@st.cache_resource
def load_processor():
class VideoProcessor:
def __init__(self):
self.caption_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
self.caption_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large")
self.vqa_processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
self.vqa_model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
# Check for available devices
if torch.cuda.is_available():
self.device = "cuda"
elif torch.backends.mps.is_available():
self.device = "mps"
else:
self.device = "cpu"
self.caption_model.to(self.device)
self.vqa_model.to(self.device)
def generate_caption(self, image):
inputs = self.caption_processor(images=image, return_tensors="pt").to(self.device)
output = self.caption_model.generate(**inputs, max_new_tokens=50)
return self.caption_processor.decode(output[0], skip_special_tokens=True)
def answer_question(self, image, question):
inputs = self.vqa_processor(image, question, return_tensors="pt").to(self.device)
outputs = self.vqa_model(**inputs)
logits = outputs.logits
idx = logits.argmax(-1).item()
return self.vqa_model.config.id2label[idx]
return VideoProcessor()
def get_video_source(source_type, source_path=None):
if source_type == "Webcam":
return cv2.VideoCapture(0)
elif source_type == "Video File" and source_path:
# Create a temporary file with a specific extension
temp_dir = tempfile.gettempdir()
temp_path = os.path.join(temp_dir, 'temp_video.mp4')
with open(temp_path, 'wb') as f:
f.write(source_path.getvalue())
cap = cv2.VideoCapture(temp_path)
if not cap.isOpened():
st.error("Error: Could not open video file. Please ensure it's a supported format (MP4 with H.264 encoding recommended)")
return None
return cap
elif source_type == "RTSP Stream" and source_path:
return cv2.VideoCapture(source_path)
return None
def process_video(stop_event, frame_queue, caption_queue, processor, source_type, source_path=None):
cap = get_video_source(source_type, source_path)
last_caption_time = time.time()
while not stop_event.is_set():
ret, frame = cap.read()
if not ret:
break
frame = cv2.resize(frame, (800, 600))
current_time = time.time()
# Generate caption every 8 seconds
if current_time - last_caption_time >= 8.0:
img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
caption = processor.generate_caption(img)
timestamp = datetime.now().strftime("%H:%M:%S")
try:
if caption_queue.full():
caption_queue.get_nowait()
caption_queue.put_nowait({'timestamp': timestamp, 'caption': caption})
last_caption_time = current_time
except queue.Full:
pass
try:
if frame_queue.full():
frame_queue.get_nowait()
frame_queue.put_nowait(frame)
except queue.Full:
pass
# time.sleep(0.03)
cap.release()
def main():
initialize_state()
# Main title
st.title("Securade.ai Sentinel")
# Create three columns for layout
video_col, caption_col, qa_col = st.columns([0.4, 0.3, 0.3])
# Video column
with video_col:
st.subheader("Video Feed")
# Video source selection
source_type = "Video File"
source_path = None
uploaded_file = None
if source_type == "Video File":
uploaded_file = st.file_uploader("Choose a video file", type=['mp4', 'avi', 'mov'])
if uploaded_file:
source_path = BytesIO(uploaded_file.getvalue())
elif source_type == "RTSP Stream":
source_path = st.text_input("Enter RTSP URL", placeholder="rtsp://your-camera-url")
start_stop = st.button(
"Start Surveillance" if not st.session_state.is_streaming else "Stop Surveillance"
)
video_placeholder = st.empty()
if start_stop:
if not st.session_state.is_streaming:
# Start surveillance
if st.session_state.processor is None:
st.session_state.processor = load_processor()
st.session_state.stop_event.clear()
st.session_state.frame_queue = queue.Queue(maxsize=1)
st.session_state.caption_queue = queue.Queue(maxsize=10)
st.session_state.thread = threading.Thread(
target=process_video,
args=(
st.session_state.stop_event,
st.session_state.frame_queue,
st.session_state.caption_queue,
st.session_state.processor,
source_type,
source_path
),
daemon=True
)
st.session_state.thread.start()
st.session_state.is_streaming = True
else:
# Stop surveillance
st.session_state.stop_event.set()
if st.session_state.thread:
st.session_state.thread.join(timeout=1.0)
st.session_state.frame = None
st.session_state.is_streaming = False
video_placeholder.empty()
# Caption column
with caption_col:
st.subheader("Scene Analysis")
caption_placeholder = st.empty()
# Q&A column
with qa_col:
st.subheader("Visual Q&A")
question = st.text_input("Ask a question about the scene:")
ask_button = st.button("Ask")
answer_placeholder = st.empty()
if ask_button and question and st.session_state.frame is not None:
img = Image.fromarray(cv2.cvtColor(st.session_state.frame, cv2.COLOR_BGR2RGB))
answer = st.session_state.processor.answer_question(img, question)
answer_placeholder.markdown(f"**Answer:** {answer}")
# Update loop
if st.session_state.is_streaming:
placeholder = st.empty()
while True:
try:
# Update video frame
frame = st.session_state.frame_queue.get_nowait()
st.session_state.frame = frame
video_placeholder.image(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
# Update captions
while not st.session_state.caption_queue.empty():
new_caption = st.session_state.caption_queue.get_nowait()
st.session_state.captions.append(new_caption)
st.session_state.captions = st.session_state.captions[-5:] # Keep last 5 captions
if st.session_state.captions:
caption_text = "\n\n".join([
f"**[{cap['timestamp']}]** {cap['caption']}"
for cap in reversed(st.session_state.captions)
])
caption_placeholder.markdown(caption_text)
except queue.Empty:
# time.sleep(0.01)
continue
if __name__ == "__main__":
main()