processing-image / src /streamlit_app.py
fouadmahmoud283-ai
fixing camera n1235
8856829
Raw
History Blame Contribute Delete
38.6 kB
import streamlit as st
import torch
import cv2
import numpy as np
from PIL import Image
import io
import time
import plotly.graph_objects as go
import plotly.express as px
import pandas as pd
from pathlib import Path
import tempfile
import os
import logging
import warnings
import requests
import threading
import av
import imageio.v2 as imageio
from streamlit_webrtc import webrtc_streamer, VideoProcessorBase
# Suppress WebRTC/asyncio warnings and errors
logging.getLogger('aioice').setLevel(logging.CRITICAL)
logging.getLogger('asyncio').setLevel(logging.CRITICAL)
logging.getLogger('streamlit_webrtc').setLevel(logging.ERROR)
warnings.filterwarnings('ignore', category=DeprecationWarning)
warnings.filterwarnings('ignore', category=FutureWarning)
# Page config
st.set_page_config(
page_title="🦽 AI Wheelchair Navigation System",
page_icon="🦽",
layout="wide",
initial_sidebar_state="expanded",
menu_items={
'Get Help': 'https://huggingface.co/spaces/fouadmahmoud281/processing-image',
'About': 'AI Wheelchair Navigation System - Graduation Project'
}
)
# Custom CSS for wheelchair theme
st.markdown("""
<style>
.main-header {
background: linear-gradient(90deg, #2E86AB 0%, #A23B72 50%, #F18F01 100%);
color: white;
padding: 2rem;
border-radius: 10px;
margin-bottom: 2rem;
text-align: center;
}
.metric-card {
background: #f8f9fa;
padding: 1rem;
border-radius: 10px;
border-left: 4px solid #2E86AB;
margin: 0.5rem 0;
}
.detection-box {
background: #e8f4f8;
padding: 1rem;
border-radius: 8px;
border: 1px solid #2E86AB;
margin: 0.5rem 0;
}
.safety-alert {
background: #fff3cd;
color: #856404;
padding: 1rem;
border-radius: 8px;
border-left: 4px solid #ffc107;
margin: 1rem 0;
}
.success-alert {
background: #d4edda;
color: #155724;
padding: 1rem;
border-radius: 8px;
border-left: 4px solid #28a745;
margin: 1rem 0;
}
.wheelchair-icon {
font-size: 2rem;
color: #2E86AB;
}
.stButton > button {
background: linear-gradient(90deg, #2E86AB, #A23B72);
color: white;
border-radius: 20px;
border: none;
padding: 0.5rem 2rem;
font-weight: bold;
}
.stSelectbox > div > div {
border-radius: 10px;
}
</style>
""", unsafe_allow_html=True)
# Initialize session state
if 'model' not in st.session_state:
st.session_state.model = None
if 'detection_history' not in st.session_state:
st.session_state.detection_history = []
if 'current_image' not in st.session_state:
st.session_state.current_image = None
if 'demo_selected' not in st.session_state:
st.session_state.demo_selected = False
if 'processed_video_path' not in st.session_state:
st.session_state.processed_video_path = None
if 'video_stats' not in st.session_state:
st.session_state.video_stats = None
if 'processed_video_mime' not in st.session_state:
st.session_state.processed_video_mime = None
# Demo images URLs (publicly accessible images)
DEMO_IMAGES = {
"Indoor Scene": "https://plus.unsplash.com/premium_photo-1661346079168-5152a94fee62?q=80&w=869&auto=format&fit=crop&ixlib=rb-4.1.0&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D", # Office/indoor
"Outdoor Scene": "https://images.unsplash.com/photo-1767034241658-0319e450c0eb?q=80&w=870&auto=format&fit=crop&ixlib=rb-4.1.0&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D", # Street scene
"Crowded Area": "https://images.unsplash.com/photo-1723930298143-48843627a478?q=80&w=387&auto=format&fit=crop&ixlib=rb-4.1.0&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D" # Crowded place
}
def load_demo_image(demo_name):
"""Load demo image from URL"""
try:
url = DEMO_IMAGES[demo_name]
response = requests.get(url, timeout=10)
response.raise_for_status()
image = Image.open(io.BytesIO(response.content))
return image
except Exception as e:
st.error(f"Error loading demo image: {e}")
return None
# Wheelchair-relevant classes from COCO
WHEELCHAIR_CLASSES = {
0: 'person', # People to avoid/navigate around
1: 'bicycle', # Other mobility devices
2: 'car', # Vehicles to avoid
3: 'motorcycle', # Vehicles to avoid
5: 'bus', # Large vehicles
7: 'truck', # Large vehicles
9: 'traffic light', # Navigation signals
11: 'stop sign', # Navigation signals
24: 'backpack', # Personal items/obstacles
26: 'handbag', # Personal items/obstacles
56: 'chair', # Furniture/obstacles
58: 'potted plant' # Environmental obstacles
}
CLASS_COLORS = {
'person': '#FF6B6B',
'bicycle': '#4ECDC4',
'car': '#45B7D1',
'motorcycle': '#96CEB4',
'bus': '#FECA57',
'truck': '#FF9F43',
'traffic light': '#6C5CE7',
'stop sign': '#FD79A8',
'chair': '#A0E7E5',
'backpack': '#DDA0DD',
'handbag': '#F7DC6F',
'potted plant': '#82E0AA'
}
@st.cache_resource
def load_model(confidence_threshold=0.5):
"""Load YOLOv5 model with caching"""
try:
# Try to load from local path first (for development)
model_paths = [
"wheelchair_runs/wheelchair_exp5/weights/best.pt",
"wheelchair_runs/wheelchair_exp/weights/best.pt",
"best.pt" # Fallback
]
model = None
for path in model_paths:
if os.path.exists(path):
model = torch.hub.load('ultralytics/yolov5', 'custom', path=path, trust_repo=True)
break
# If no local model, load pretrained YOLOv5s
if model is None:
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True, trust_repo=True)
model.conf = confidence_threshold
model.iou = 0.45
return model
except Exception as e:
st.error(f"Error loading model: {e}")
return None
def get_navigation_advice(detections, image_width=640):
"""Generate navigation advice based on detections"""
advice = []
safety_level = "🟒 SAFE"
if detections is None or len(detections) == 0:
return ["βœ… Clear path ahead"], safety_level
critical_detections = 0
for _, detection in detections.iterrows():
class_id = int(detection['class'])
class_name = detection['name']
confidence = detection['confidence']
x_center = (detection['xmin'] + detection['xmax']) / 2
y_center = (detection['ymin'] + detection['ymax']) / 2
# Determine position relative to wheelchair
if x_center < image_width * 0.33:
position = "left"
elif x_center > image_width * 0.67:
position = "right"
else:
position = "ahead"
critical_detections += 1
# Generate specific advice based on object type and position
if class_name == 'person':
if position == "ahead":
advice.append(f"⚠️ PERSON DETECTED AHEAD - STOP and wait for clear path")
safety_level = "πŸ”΄ CRITICAL"
else:
advice.append(f"πŸ‘€ Person on {position} (conf: {confidence:.0%})")
if safety_level == "🟒 SAFE":
safety_level = "🟑 CAUTION"
elif class_name in ['car', 'truck', 'bus', 'motorcycle']:
advice.append(f"πŸš— {class_name.title()} {position} - proceed with extreme caution")
safety_level = "πŸ”΄ CRITICAL" if position == "ahead" else "🟑 CAUTION"
elif class_name in ['traffic light', 'stop sign']:
advice.append(f"🚦 {class_name.replace('_', ' ').title()} detected - follow traffic rules")
elif class_name == 'chair':
advice.append(f"πŸͺ‘ Chair detected on {position} - navigate around")
if position == "ahead":
safety_level = "🟑 CAUTION"
elif class_name in ['backpack', 'handbag']:
advice.append(f"πŸŽ’ Personal item on {position} - person nearby")
elif class_name == 'potted plant':
advice.append(f"πŸͺ΄ Obstacle on {position} - adjust path")
if critical_detections > 2:
safety_level = "πŸ”΄ CRITICAL"
advice.insert(0, "⚠️ MULTIPLE OBSTACLES AHEAD - STOP AND REASSESS")
return advice[:5], safety_level # Limit to top 5 pieces of advice
def process_image(image, model, conf_threshold):
"""Process image and return results"""
if model is None:
return None, None, []
# Run inference
results = model(image)
# Get detections
detections = results.pandas().xyxy[0]
# Filter to wheelchair-relevant classes
relevant_detections = detections[detections['name'].isin(WHEELCHAIR_CLASSES.values())]
# Get rendered image
rendered_img = results.render()[0]
rendered_img = cv2.cvtColor(rendered_img, cv2.COLOR_BGR2RGB)
return rendered_img, relevant_detections, results
def process_video(video_path, model, conf_threshold, frame_skip=1):
"""Process video and return output path + stats."""
if model is None:
return None, None
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
return None, None
fps = cap.get(cv2.CAP_PROP_FPS) or 24
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
output_fd, output_path = tempfile.mkstemp(suffix=".mp4")
os.close(output_fd)
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 0
processed_frames = 0
detection_counts = {}
progress = st.progress(0, text="Processing video frames...")
container = None
stream = None
codec_name = None
error_message = None
output_mime = "video/mp4"
def _process_frame(frame_bgr):
nonlocal processed_frames
img_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
model.conf = conf_threshold
results = model(img_rgb)
detections = results.pandas().xyxy[0]
relevant = detections[detections['name'].isin(WHEELCHAIR_CLASSES.values())]
for name in relevant['name'].tolist():
detection_counts[name] = detection_counts.get(name, 0) + 1
rendered_frame = results.render()[0]
if rendered_frame.shape[1] != width or rendered_frame.shape[0] != height:
rendered_frame = cv2.resize(rendered_frame, (width, height))
processed_frames += 1
if total_frames > 0:
progress.progress(min(processed_frames / total_frames, 1.0))
return rendered_frame
def _reset_stats():
nonlocal processed_frames, detection_counts
processed_frames = 0
detection_counts = {}
progress.progress(0)
def _run_loop(write_frame):
cap.set(cv2.CAP_PROP_POS_FRAMES, 0)
frame_index = 0
while True:
ret, frame = cap.read()
if not ret:
break
frame_index += 1
if frame_skip > 1 and frame_index % frame_skip != 0:
continue
rendered_frame = _process_frame(frame)
write_frame(rendered_frame)
try:
container = av.open(output_path, mode="w")
try:
stream = container.add_stream("libx264", rate=fps)
codec_name = "libx264"
output_mime = "video/mp4"
except av.AVError:
container.close()
output_path = output_path.replace(".mp4", ".webm")
container = av.open(output_path, mode="w")
stream = container.add_stream("libvpx", rate=fps)
codec_name = "libvpx"
output_mime = "video/webm"
stream.width = width
stream.height = height
stream.pix_fmt = "yuv420p"
_run_loop(lambda rendered_frame: [container.mux(p) for p in stream.encode(av.VideoFrame.from_ndarray(rendered_frame, format="bgr24"))])
for packet in stream.encode():
container.mux(packet)
except Exception as e:
error_message = f"PyAV encode failed: {e}"
finally:
if container is not None:
container.close()
if error_message:
_reset_stats()
try:
imageio_path = output_path.replace(".webm", ".mp4")
writer = imageio.get_writer(
imageio_path,
fps=fps,
codec="libx264",
ffmpeg_params=["-pix_fmt", "yuv420p"],
format="FFMPEG"
)
_run_loop(lambda rendered_frame: writer.append_data(cv2.cvtColor(rendered_frame, cv2.COLOR_BGR2RGB)))
writer.close()
output_path = imageio_path
codec_name = "libx264"
output_mime = "video/mp4"
error_message = None
except Exception as e:
error_message = f"ImageIO H.264 encode failed: {e}"
if error_message:
_reset_stats()
try:
imageio_path = output_path.replace(".mp4", ".webm")
writer = imageio.get_writer(
imageio_path,
fps=fps,
codec="libvpx",
ffmpeg_params=["-pix_fmt", "yuv420p"],
format="FFMPEG"
)
_run_loop(lambda rendered_frame: writer.append_data(cv2.cvtColor(rendered_frame, cv2.COLOR_BGR2RGB)))
writer.close()
output_path = imageio_path
codec_name = "libvpx"
output_mime = "video/webm"
error_message = None
except Exception as e:
error_message = f"ImageIO WebM encode failed: {e}"
if error_message:
_reset_stats()
try:
fourcc = cv2.VideoWriter_fourcc(*"mp4v")
writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
_run_loop(lambda rendered_frame: writer.write(rendered_frame))
writer.release()
error_message = None
codec_name = "mp4v"
output_mime = "video/mp4"
except Exception as e:
error_message = f"OpenCV encode failed: {e}"
output_path = None
progress.empty()
cap.release()
stats = {
"total_frames": total_frames,
"processed_frames": processed_frames,
"detection_counts": detection_counts,
"codec": codec_name,
"error": error_message,
"mime": output_mime
}
return output_path, stats
def create_detection_chart(detections):
"""Create a bar chart of detections"""
if detections is None or len(detections) == 0:
return None
detection_counts = detections['name'].value_counts()
fig = px.bar(
x=detection_counts.index,
y=detection_counts.values,
color=detection_counts.index,
color_discrete_map=CLASS_COLORS,
title="πŸ“Š Detected Objects Count",
labels={'x': 'Object Type', 'y': 'Count'}
)
fig.update_layout(
showlegend=False,
plot_bgcolor='rgba(0,0,0,0)',
paper_bgcolor='rgba(0,0,0,0)',
font=dict(size=12),
title_font=dict(size=16, color='#2E86AB')
)
return fig
def create_confidence_chart(detections):
"""Create a confidence score visualization"""
if detections is None or len(detections) == 0:
return None
fig = px.scatter(
detections,
x='name',
y='confidence',
size='confidence',
color='name',
color_discrete_map=CLASS_COLORS,
title="🎯 Detection Confidence Scores",
labels={'confidence': 'Confidence Score', 'name': 'Object Type'}
)
fig.update_layout(
showlegend=False,
plot_bgcolor='rgba(0,0,0,0)',
paper_bgcolor='rgba(0,0,0,0)',
font=dict(size=12),
title_font=dict(size=16, color='#2E86AB')
)
return fig
def get_rtc_configuration():
"""Build RTC config from env vars (TURN optional)."""
turn_url = os.getenv("TURN_URL")
turn_username = os.getenv("TURN_USERNAME")
turn_password = os.getenv("TURN_PASSWORD")
ice_servers = [
{"urls": ["stun:stun.l.google.com:19302", "stun:stun1.l.google.com:19302"]}
]
if turn_url and turn_username and turn_password:
ice_servers.append({
"urls": [turn_url],
"username": turn_username,
"credential": turn_password
})
return {"iceServers": ice_servers}
class RealtimeVideoProcessor(VideoProcessorBase):
def __init__(self, model, conf_threshold):
self.model = model
self.conf_threshold = conf_threshold
self.last_detections = None
self.lock = threading.Lock()
def recv(self, frame):
img_bgr = frame.to_ndarray(format="bgr24")
img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
if self.model is None:
return frame
self.model.conf = self.conf_threshold
results = self.model(img_rgb)
detections = results.pandas().xyxy[0]
relevant_detections = detections[detections['name'].isin(WHEELCHAIR_CLASSES.values())]
with self.lock:
self.last_detections = relevant_detections
rendered_img = results.render()[0]
return av.VideoFrame.from_ndarray(rendered_img, format="bgr24")
# Main app
def main():
# Header
st.markdown("""
<div class="main-header">
<h1>🦽 AI Wheelchair Navigation System</h1>
<p>Intelligent Computer Vision for Safe Wheelchair Navigation</p>
<p><strong>Graduation Project</strong> | YOLOv5 Object Detection | Raspberry Pi Optimized</p>
</div>
""", unsafe_allow_html=True)
# Sidebar
with st.sidebar:
st.markdown("### πŸ”§ Model Configuration")
# Model settings
conf_threshold = st.slider(
"Confidence Threshold",
min_value=0.1,
max_value=1.0,
value=0.5,
step=0.1,
help="Minimum confidence for object detection"
)
st.markdown("### πŸ“Š Model Performance")
st.markdown("""
<div class="metric-card">
<strong>🎯 Overall Accuracy:</strong> 87.5% mAP@0.5<br>
<strong>πŸ‘€ Person Detection:</strong> 84.4%<br>
<strong>πŸͺ‘ Chair Detection:</strong> 93.9%<br>
<strong>πŸš— Vehicle Detection:</strong> 56.3%<br>
<strong>βš–οΈ Model Size:</strong> 14.7 MB
</div>
""", unsafe_allow_html=True)
st.markdown("### 🎯 Wheelchair-Relevant Objects")
for class_id, class_name in WHEELCHAIR_CLASSES.items():
color = CLASS_COLORS.get(class_name, '#808080')
st.markdown(f'<span style="color: {color}; font-weight: bold;">●</span> {class_name.title()}', unsafe_allow_html=True)
# Main content tabs
tab1, tab2, tab3, tab4, tab5 = st.tabs(["πŸ“Έ Live Detection", "πŸ“Ή Realtime Camera", "πŸ“Š Analytics", "ℹ️ About", "πŸš€ Deployment"])
with tab1:
col1, col2 = st.columns([2, 1])
with col1:
st.markdown("### πŸ“· Upload Image for Detection")
uploaded_file = st.file_uploader(
"Choose an image...",
type=['jpg', 'jpeg', 'png', 'bmp'],
help="Upload an image to test wheelchair navigation detection",
key="image_uploader"
)
st.markdown("### πŸŽ₯ Upload Video for Detection")
uploaded_video = st.file_uploader(
"Choose a video...",
type=['mp4', 'avi', 'mov', 'mkv'],
help="Upload a video to run obstacle detection",
key="video_uploader"
)
# Demo images buttons
st.markdown("### 🎬 Or Try Demo Images")
demo_col1, demo_col2, demo_col3 = st.columns(3)
demo_image = None
with demo_col1:
if st.button("πŸͺ Indoor Scene", use_container_width=True):
st.session_state.demo_selected = True
demo_image = load_demo_image("Indoor Scene")
if demo_image:
st.session_state.current_image = demo_image
with demo_col2:
if st.button("πŸ™οΈ Outdoor Scene", use_container_width=True):
st.session_state.demo_selected = True
demo_image = load_demo_image("Outdoor Scene")
if demo_image:
st.session_state.current_image = demo_image
with demo_col3:
if st.button("🚢 Crowded Area", use_container_width=True):
st.session_state.demo_selected = True
demo_image = load_demo_image("Crowded Area")
if demo_image:
st.session_state.current_image = demo_image
with col2:
st.markdown("### πŸ›‘οΈ Safety Status")
safety_placeholder = st.empty()
st.markdown("### 🧭 Navigation Advice")
advice_placeholder = st.empty()
# Process uploaded or demo image
image = None
if uploaded_file is not None:
st.session_state.demo_selected = False
image = Image.open(uploaded_file)
st.session_state.current_image = image
elif st.session_state.demo_selected and st.session_state.current_image is not None:
image = st.session_state.current_image
# Process uploaded video
if uploaded_video is not None:
st.session_state.demo_selected = False
if st.session_state.model is None:
with st.spinner("πŸ€– Loading AI model..."):
st.session_state.model = load_model(conf_threshold)
if st.button("▢️ Run Video Detection", use_container_width=True):
with tempfile.NamedTemporaryFile(delete=False, suffix=Path(uploaded_video.name).suffix) as tmp:
tmp.write(uploaded_video.getbuffer())
tmp_path = tmp.name
with st.spinner("πŸ” Processing video for obstacles..."):
output_path, stats = process_video(tmp_path, st.session_state.model, conf_threshold, frame_skip=1)
st.session_state.processed_video_path = output_path
st.session_state.video_stats = stats
st.session_state.processed_video_mime = stats.get("mime") if stats else None
try:
os.remove(tmp_path)
except Exception:
pass
if st.session_state.processed_video_path:
st.markdown("### 🎬 Processed Video")
try:
if os.path.exists(st.session_state.processed_video_path) and os.path.getsize(st.session_state.processed_video_path) > 0:
with open(st.session_state.processed_video_path, "rb") as f:
st.video(f.read(), format=st.session_state.processed_video_mime or "video/mp4")
else:
st.error("Processed video file is empty or missing.")
except Exception as e:
st.error(f"Unable to load processed video: {e}")
if st.session_state.video_stats:
st.markdown("### πŸ“Š Video Detection Summary")
stats = st.session_state.video_stats
st.metric("🎞️ Total Frames", stats.get("total_frames", 0))
st.metric("βœ… Processed Frames", stats.get("processed_frames", 0))
if stats.get("codec"):
st.caption(f"Codec: {stats.get('codec')}")
if stats.get("mime"):
st.caption(f"MIME: {stats.get('mime')}")
if stats.get("error"):
st.error(stats.get("error"))
if stats.get("detection_counts"):
counts_df = pd.DataFrame(
sorted(stats["detection_counts"].items(), key=lambda x: x[1], reverse=True),
columns=["Object", "Count"]
)
st.dataframe(counts_df, use_container_width=True)
if image is not None:
col1, col2 = st.columns(2)
with col1:
st.markdown("#### πŸ“Έ Original Image")
st.image(image, width=500)
# Load model
if st.session_state.model is None:
with st.spinner("πŸ€– Loading AI model..."):
st.session_state.model = load_model(conf_threshold)
if st.session_state.model is not None:
# Process image
with st.spinner("πŸ” Analyzing image for obstacles..."):
start_time = time.time()
rendered_img, detections, results = process_image(image, st.session_state.model, conf_threshold)
processing_time = time.time() - start_time
with col2:
st.markdown("#### 🎯 Detection Results")
if rendered_img is not None:
st.image(rendered_img, width=500)
# Generate navigation advice
advice, safety_level = get_navigation_advice(detections, image.width)
# Update safety status
with safety_placeholder.container():
if "CRITICAL" in safety_level:
st.markdown(f'<div class="safety-alert"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
elif "CAUTION" in safety_level:
st.markdown(f'<div style="background: #fff3cd; color: #856404; padding: 1rem; border-radius: 8px; border-left: 4px solid #ffc107;"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
else:
st.markdown(f'<div class="success-alert"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
# Display navigation advice
with advice_placeholder.container():
for advice_text in advice:
st.markdown(f'<div class="detection-box">{advice_text}</div>', unsafe_allow_html=True)
# Performance metrics
st.markdown("### ⚑ Performance Metrics")
perf_col1, perf_col2, perf_col3, perf_col4 = st.columns(4)
with perf_col1:
st.metric("⏱️ Processing Time", f"{processing_time:.2f}s")
with perf_col2:
fps = 1 / processing_time if processing_time > 0 else 0
st.metric("🎬 Estimated FPS", f"{fps:.1f}")
with perf_col3:
total_detections = len(detections) if detections is not None else 0
st.metric("πŸ” Objects Detected", total_detections)
with perf_col4:
relevant_count = len(detections) if detections is not None else 0
st.metric("🎯 Relevant Objects", relevant_count)
# Detailed detection results
if detections is not None and len(detections) > 0:
st.markdown("### πŸ“‹ Detailed Detection Results")
# Create a formatted dataframe
display_df = detections[['name', 'confidence', 'xmin', 'ymin', 'xmax', 'ymax']].copy()
display_df['confidence'] = display_df['confidence'].apply(lambda x: f"{x:.1%}")
display_df.columns = ['Object', 'Confidence', 'X Min', 'Y Min', 'X Max', 'Y Max']
st.dataframe(display_df, use_container_width=True)
with tab2:
st.markdown("### πŸ“Ή Realtime Camera Detection")
st.caption("Start the camera to run live detection on each frame.")
if 'model' not in st.session_state:
st.session_state.model = None
if st.session_state.model is None:
with st.spinner("πŸ€– Loading AI model..."):
st.session_state.model = load_model(conf_threshold)
model = st.session_state.model
rtc_config = get_rtc_configuration()
if not os.getenv("TURN_URL"):
st.info("Live camera may fail on some networks without TURN. Set TURN_URL, TURN_USERNAME, TURN_PASSWORD to improve connectivity.")
webrtc_ctx = webrtc_streamer(
key="realtime_camera",
video_processor_factory=lambda: RealtimeVideoProcessor(model, conf_threshold),
media_stream_constraints={"video": True, "audio": False},
async_processing=True,
rtc_configuration=rtc_config
)
realtime_detections = None
if webrtc_ctx.video_processor:
with webrtc_ctx.video_processor.lock:
realtime_detections = webrtc_ctx.video_processor.last_detections
if realtime_detections is not None:
st.markdown("### πŸ›‘οΈ Safety Status")
advice, safety_level = get_navigation_advice(realtime_detections, image_width=640)
if "CRITICAL" in safety_level:
st.markdown(f'<div class="safety-alert"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
elif "CAUTION" in safety_level:
st.markdown(f'<div style="background: #fff3cd; color: #856404; padding: 1rem; border-radius: 8px; border-left: 4px solid #ffc107;"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
else:
st.markdown(f'<div class="success-alert"><strong>{safety_level}</strong></div>', unsafe_allow_html=True)
st.markdown("### 🧭 Navigation Advice")
for advice_text in advice:
st.markdown(f'<div class="detection-box">{advice_text}</div>', unsafe_allow_html=True)
if len(realtime_detections) > 0:
st.markdown("### πŸ“‹ Latest Detections")
display_df = realtime_detections[['name', 'confidence', 'xmin', 'ymin', 'xmax', 'ymax']].copy()
display_df['confidence'] = display_df['confidence'].apply(lambda x: f"{x:.1%}")
display_df.columns = ['Object', 'Confidence', 'X Min', 'Y Min', 'X Max', 'Y Max']
st.dataframe(display_df, use_container_width=True)
with tab3:
st.markdown("### πŸ“Š Detection Analytics")
if uploaded_file is not None and 'detections' in locals() and detections is not None:
col1, col2 = st.columns(2)
with col1:
chart1 = create_detection_chart(detections)
if chart1:
st.plotly_chart(chart1, use_container_width=True)
with col2:
chart2 = create_confidence_chart(detections)
if chart2:
st.plotly_chart(chart2, use_container_width=True)
# Detection statistics
st.markdown("### πŸ“ˆ Statistics")
if len(detections) > 0:
stats_col1, stats_col2, stats_col3 = st.columns(3)
with stats_col1:
avg_conf = detections['confidence'].mean()
st.metric("πŸ“Š Average Confidence", f"{avg_conf:.1%}")
with stats_col2:
max_conf = detections['confidence'].max()
st.metric("🎯 Highest Confidence", f"{max_conf:.1%}")
with stats_col3:
unique_classes = detections['name'].nunique()
st.metric("🏷️ Unique Object Types", unique_classes)
else:
st.info("πŸ“Έ Upload an image in the 'Live Detection' tab to see analytics")
with tab4:
st.markdown("### ℹ️ About This System")
col1, col2 = st.columns([2, 1])
with col1:
st.markdown("""
#### 🦽 AI Wheelchair Navigation System
This intelligent computer vision system is designed to assist wheelchair users with safe navigation by detecting and identifying potential obstacles, people, vehicles, and navigation signals in real-time.
**🎯 Key Features:**
- **Real-time Object Detection**: Identifies 12 wheelchair-relevant object types
- **Safety Warnings**: Provides immediate alerts for potential hazards
- **Navigation Guidance**: Offers contextual advice for safe path planning
- **Raspberry Pi Optimized**: Lightweight model for edge deployment
- **High Accuracy**: 87.5% mAP@0.5 overall accuracy
**πŸ”§ Technical Specifications:**
- **Model**: YOLOv5s (Small) - optimized for speed and accuracy
- **Input Size**: 416x416 pixels
- **Model Size**: 14.7 MB (perfect for embedded systems)
- **Target Platform**: Raspberry Pi 4
- **Processing Speed**: 5-10 FPS on Raspberry Pi
**πŸŽ“ Graduation Project Context:**
This system represents a comprehensive computer vision solution for assistive technology, demonstrating:
- Advanced deep learning implementation
- Edge computing optimization
- Real-world application development
- Safety-critical system design
""")
with col2:
st.markdown("""
#### πŸ† Model Performance
**Overall Metrics:**
- mAP@0.5: 87.5%
- mAP@0.5:0.95: 63.6%
- Precision: 88.6%
- Recall: 80.8%
**Class-Specific Performance:**
- Person: 84.4% mAP
- Chair: 93.9% mAP
- Vehicle: 56.3% mAP
- Bicycle: 80.9% mAP
**πŸ”’ Safety Features:**
- Emergency obstacle detection
- Multi-level alert system
- Contextual navigation advice
- Real-time processing
""")
st.markdown("### πŸ› οΈ Technology Stack")
tech_col1, tech_col2, tech_col3, tech_col4 = st.columns(4)
with tech_col1:
st.markdown("""
**🧠 AI/ML**
- YOLOv5
- PyTorch
- OpenCV
- NumPy
""")
with tech_col2:
st.markdown("""
**🌐 Web App**
- Streamlit
- Plotly
- PIL/Pillow
- Pandas
""")
with tech_col3:
st.markdown("""
**⚑ Deployment**
- Hugging Face Spaces
- Docker
- Git LFS
- ONNX (optional)
""")
with tech_col4:
st.markdown("""
**πŸ”§ Hardware**
- Raspberry Pi 4
- USB/Pi Camera
- MicroSD Storage
- Power Supply
""")
with tab5:
st.markdown("### πŸš€ Deployment Information")
col1, col2 = st.columns(2)
with col1:
st.markdown("""
#### πŸ“¦ Hugging Face Spaces Deployment
This application is deployed on Hugging Face Spaces, providing:
- **Free hosting** for demonstration purposes
- **Easy sharing** with project evaluators
- **Scalable infrastructure** for multiple users
- **Integrated CI/CD** for automatic updates
**πŸ”— Deployment Features:**
- Real-time inference on uploaded images
- Interactive web interface
- Performance analytics and visualization
- Mobile-responsive design
""")
with col2:
st.markdown("""
#### 🏠 Local/Raspberry Pi Deployment
For real wheelchair deployment:
1. **Download the deployment package**
2. **Transfer to Raspberry Pi**
3. **Install dependencies**
4. **Connect camera**
5. **Run inference script**
**πŸ“‹ Requirements:**
- Raspberry Pi 4 (4GB RAM recommended)
- Python 3.7+
- PyTorch (CPU version)
- USB Camera or Pi Camera
""")
st.markdown("### πŸ’» Code Repository")
st.markdown("""
**πŸ“ Project Structure:**
```
wheelchair_deployment/
β”œβ”€β”€ wheelchair_model.pt # Trained model weights
β”œβ”€β”€ wheelchair_inference.py # Raspberry Pi inference script
β”œβ”€β”€ requirements_rpi.txt # Dependencies
β”œβ”€β”€ README_deployment.md # Setup instructions
└── wheelchair_config.yaml # Model configuration
```
""")
st.markdown("### 🀝 Integration Guidelines")
st.markdown("""
**For Wheelchair Integration:**
1. **Motor Control Interface**: Connect detection results to wheelchair motor control system
2. **Safety Protocols**: Implement emergency stop and collision avoidance
3. **User Interface**: Add audio/visual feedback for navigation guidance
4. **Sensor Fusion**: Combine with ultrasonic/LiDAR sensors for enhanced safety
5. **Custom Training**: Collect and label wheelchair-specific navigation data
""")
if __name__ == "__main__":
main()