visual-odometry / app.py
Tremick's picture
Upload 3 files
f1619b7 verified
"""
Optical Flow Server for Hugging Face Spaces
============================================
Complete server for cloud-based optical flow processing.
"""
import gradio as gr
import cv2
import numpy as np
import base64
from io import BytesIO
from PIL import Image
from sklearn.cluster import DBSCAN
import colorsys
from collections import deque
import uuid
import time
import json
class ServerOpticalFlowEngine:
"""Optical Flow Engine for server-side processing with persistent tracking."""
def __init__(self):
self.prev_gray = None
self.trail_points = {}
self.max_trail_length = 30
self.bg_subtractor = cv2.createBackgroundSubtractorMOG2(
history=500, varThreshold=16, detectShadows=True
)
self.color_pool = self._generate_colors(20)
# Persistent object tracking
self.tracked_objects = {}
self.next_object_id = 1
self.max_frames_missing = 15
def _generate_colors(self, n):
colors = []
for i in range(n):
hue = i / n
rgb = colorsys.hsv_to_rgb(hue, 0.9, 0.95)
bgr = (int(rgb[2] * 255), int(rgb[1] * 255), int(rgb[0] * 255))
colors.append(bgr)
return colors
def compute_flow(self, gray):
if self.prev_gray is None:
self.prev_gray = gray.copy()
return None
flow = cv2.calcOpticalFlowFarneback(
self.prev_gray, gray, None,
pyr_scale=0.5, levels=5, winsize=13,
iterations=10, poly_n=5, poly_sigma=1.1,
flags=cv2.OPTFLOW_FARNEBACK_GAUSSIAN
)
self.prev_gray = gray.copy()
return flow
def segment_motion(self, frame):
fg_mask = self.bg_subtractor.apply(frame)
fg_mask[fg_mask == 127] = 0
kernel_small = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
fg_mask = cv2.erode(fg_mask, kernel_small, iterations=2)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_OPEN, kernel)
kernel_large = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15))
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_CLOSE, kernel_large)
fg_mask = cv2.dilate(fg_mask, kernel_small, iterations=1)
return fg_mask
def compute_iou(self, bbox1, bbox2):
"""Compute IoU between two bounding boxes."""
x1_1, y1_1, x2_1, y2_1 = bbox1
x1_2, y1_2, x2_2, y2_2 = bbox2
xi1, yi1 = max(x1_1, x1_2), max(y1_1, y1_2)
xi2, yi2 = min(x2_1, x2_2), min(y2_1, y2_2)
if xi2 <= xi1 or yi2 <= yi1:
return 0.0
inter = (xi2 - xi1) * (yi2 - yi1)
area1 = (x2_1 - x1_1) * (y2_1 - y1_1)
area2 = (x2_2 - x1_2) * (y2_2 - y1_2)
return inter / (area1 + area2 - inter) if (area1 + area2 - inter) > 0 else 0.0
def match_and_track(self, detected):
"""Match detections to tracked objects using IoU."""
if not detected:
for oid in list(self.tracked_objects.keys()):
self.tracked_objects[oid]['missing'] += 1
if self.tracked_objects[oid]['missing'] > self.max_frames_missing:
del self.tracked_objects[oid]
if oid in self.trail_points:
del self.trail_points[oid]
return {}
det_list = list(detected.values())
tracked_ids = list(self.tracked_objects.keys())
matched_t, matched_d, result = set(), set(), {}
for tid in tracked_ids:
best_iou, best_idx = 0.3, None
for i, det in enumerate(det_list):
if i in matched_d:
continue
iou = self.compute_iou(self.tracked_objects[tid]['bbox'], det['bbox'])
if iou > best_iou:
best_iou, best_idx = iou, i
if best_idx is not None:
matched_t.add(tid)
matched_d.add(best_idx)
det = det_list[best_idx]
self.tracked_objects[tid].update({
'centroid': det['centroid'], 'bbox': det['bbox'],
'velocity': det['velocity'], 'area': det['area'],
'contour': det.get('contour'), 'missing': 0
})
result[tid] = self.tracked_objects[tid]
for tid in tracked_ids:
if tid not in matched_t:
self.tracked_objects[tid]['missing'] += 1
if self.tracked_objects[tid]['missing'] > self.max_frames_missing:
del self.tracked_objects[tid]
if tid in self.trail_points:
del self.trail_points[tid]
for i, det in enumerate(det_list):
if i not in matched_d:
nid = self.next_object_id
self.next_object_id += 1
self.tracked_objects[nid] = {
'centroid': det['centroid'], 'bbox': det['bbox'],
'velocity': det['velocity'], 'area': det['area'],
'contour': det.get('contour'), 'missing': 0
}
result[nid] = self.tracked_objects[nid]
return result
def detect_objects(self, mask, flow=None, min_area=2000):
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if not contours:
return None, self.match_and_track({})
label_image = np.full(mask.shape, -1, dtype=np.int32)
objects = {}
label_id = 0
for contour in contours:
area = cv2.contourArea(contour)
if area < min_area:
continue
x, y, w, h = cv2.boundingRect(contour)
aspect_ratio = float(h) / w if w > 0 else 0
if aspect_ratio < 0.2 or aspect_ratio > 8.0:
continue
hull = cv2.convexHull(contour)
hull_area = cv2.contourArea(hull)
if hull_area > 0 and (area / hull_area) < 0.3:
continue
cv2.drawContours(label_image, [contour], -1, label_id, -1)
M = cv2.moments(contour)
cx = int(M['m10'] / M['m00']) if M['m00'] > 0 else x + w // 2
cy = int(M['m01'] / M['m00']) if M['m00'] > 0 else y + h // 2
avg_velocity = (0, 0)
if flow is not None:
contour_mask = np.zeros(mask.shape, dtype=np.uint8)
cv2.drawContours(contour_mask, [contour], -1, 255, -1)
flow_x = flow[:, :, 0][contour_mask > 0]
flow_y = flow[:, :, 1][contour_mask > 0]
if len(flow_x) > 0:
avg_velocity = (float(np.mean(flow_x)), float(np.mean(flow_y)))
objects[label_id] = {
'centroid': (cx, cy), 'bbox': (x, y, x + w, y + h),
'velocity': avg_velocity, 'area': area, 'contour': contour
}
label_id += 1
# Apply persistent tracking
tracked = self.match_and_track(objects)
return label_image, tracked
def update_trails(self, objects):
for obj_id, obj in objects.items():
if obj_id not in self.trail_points:
self.trail_points[obj_id] = deque(maxlen=self.max_trail_length)
self.trail_points[obj_id].append(obj['centroid'])
def draw_results(self, frame, objects, label_image):
output = frame.copy()
for label, obj in objects.items():
color = self.color_pool[label % len(self.color_pool)]
x1, y1, x2, y2 = obj['bbox']
cv2.rectangle(output, (x1, y1), (x2, y2), color, 2)
cx, cy = obj['centroid']
cv2.circle(output, (cx, cy), 5, color, -1)
vx, vy = obj['velocity']
if np.sqrt(vx**2 + vy**2) > 1:
cv2.arrowedLine(output, (cx, cy),
(int(cx + vx * 5), int(cy + vy * 5)),
(0, 255, 255), 3, tipLength=0.3)
cv2.putText(output, f"Obj {label}", (x1, y1 - 25),
cv2.FONT_HERSHEY_SIMPLEX, 0.6, color, 2)
if label_image is not None:
mask = (label_image == label).astype(np.uint8) * 255
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cv2.drawContours(output, contours, -1, color, 2)
return output
def draw_trails(self, frame):
output = frame.copy()
for obj_id, trail in self.trail_points.items():
if len(trail) < 2:
continue
color = self.color_pool[obj_id % len(self.color_pool)]
points = np.array(list(trail), dtype=np.int32)
for i in range(1, len(points)):
cv2.line(output, tuple(points[i-1]), tuple(points[i]), color, int(1 + (i / len(points)) * 3))
cv2.circle(output, tuple(points[-1]), 6, color, -1)
return output
def compute_heatmap(self, flow):
if flow is None:
return None
magnitude = np.sqrt(flow[..., 0]**2 + flow[..., 1]**2)
normalized = cv2.normalize(magnitude, None, 0, 255, cv2.NORM_MINMAX)
return cv2.applyColorMap(normalized.astype(np.uint8), cv2.COLORMAP_HOT)
def process_frame(self, frame):
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
flow = self.compute_flow(gray)
motion_mask = self.segment_motion(frame)
results = {'num_objects': 0, 'tracked': frame.copy(),
'trails': frame.copy(), 'heatmap': np.zeros_like(frame)}
if flow is not None:
results['heatmap'] = self.compute_heatmap(flow)
label_image, objects = self.detect_objects(motion_mask, flow)
if objects:
results['num_objects'] = len(objects)
self.update_trails(objects)
results['tracked'] = self.draw_results(frame, objects, label_image)
results['trails'] = self.draw_trails(results['tracked'])
return results
def reset(self):
self.prev_gray = None
self.trail_points = {}
self.tracked_objects = {}
self.next_object_id = 1
# Session management
sessions = {}
SESSION_TIMEOUT = 300
def get_or_create_session(session_id):
current_time = time.time()
expired = [sid for sid, data in sessions.items() if current_time - data['last_access'] > SESSION_TIMEOUT]
for sid in expired:
del sessions[sid]
if session_id not in sessions:
sessions[session_id] = {'engine': ServerOpticalFlowEngine(), 'last_access': current_time}
else:
sessions[session_id]['last_access'] = current_time
return sessions[session_id]['engine']
def decode_frame(base64_data):
img_data = base64.b64decode(base64_data)
img = Image.open(BytesIO(img_data))
return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
def encode_frame(frame):
_, buffer = cv2.imencode('.jpg', frame, [cv2.IMWRITE_JPEG_QUALITY, 85])
return base64.b64encode(buffer).decode('utf-8')
# API Functions exposed via Gradio
def process_frame_api(frame_base64: str, session_id: str) -> str:
"""Process a frame and return JSON result."""
try:
engine = get_or_create_session(session_id)
frame = decode_frame(frame_base64)
results = engine.process_frame(frame)
return json.dumps({
'success': True, 'num_objects': results['num_objects'],
'tracked': encode_frame(results['tracked']),
'trails': encode_frame(results['trails']),
'heatmap': encode_frame(results['heatmap'])
})
except Exception as e:
return json.dumps({'success': False, 'error': str(e)})
def reset_session_api(session_id: str) -> str:
"""Reset session."""
if session_id in sessions:
sessions[session_id]['engine'].reset()
return json.dumps({'success': True})
def create_new_session() -> str:
"""Create new session."""
session_id = str(uuid.uuid4())
get_or_create_session(session_id)
return session_id
def test_image(image):
"""Test with uploaded image."""
if image is None:
return None
engine = ServerOpticalFlowEngine()
engine.process_frame(image)
return engine.process_frame(image)['tracked']
# Gradio UI with exposed API endpoints
with gr.Blocks(title="Optical Flow Server") as demo:
gr.Markdown("# 🎥 Optical Flow Processing Server")
gr.Markdown("Use: `python optical_flow_advanced.py --server " +
"https://tremick-visual-odometry.hf.space`")
with gr.Tab("Test"):
with gr.Row():
input_img = gr.Image(label="Upload Image", type="numpy")
output_img = gr.Image(label="Processed")
input_img.change(test_image, inputs=input_img, outputs=output_img)
with gr.Tab("API"):
gr.Markdown("### API Endpoints")
gr.Markdown("Use the functions below via Gradio Client:")
# Expose API functions
session_btn = gr.Button("Create Session")
session_out = gr.Textbox(label="Session ID")
session_btn.click(create_new_session, outputs=session_out)
with gr.Row():
frame_input = gr.Textbox(label="Frame (base64)", lines=3)
sid_input = gr.Textbox(label="Session ID")
process_btn = gr.Button("Process Frame")
result_out = gr.Textbox(label="Result (JSON)", lines=5)
process_btn.click(process_frame_api, inputs=[frame_input, sid_input], outputs=result_out)
# Also create simple API interface for programmatic access
api_interface = gr.Interface(
fn=process_frame_api,
inputs=[gr.Textbox(label="frame_base64"), gr.Textbox(label="session_id")],
outputs=gr.Textbox(label="result"),
api_name="process_frame"
)
demo = gr.TabbedInterface(
[demo, api_interface],
["UI", "API"]
)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860)