def main():
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
cl = serv.getClient()
cv2.namedWindow('RoniStream', cv2.WINDOW_AUTOSIZE)
while True:
data = serv.receiveData(cl, Roni.TYPE_RGB)
if len(data) > 0:
img = Coppa.decodeColorFrame(data)
cv2.imshow('RoniStream', img)
cv2.waitKey(1)
def main():
global keyPress
global gui
global showConf
global clients
global nindex
global connHandler
global serv
print("Reading in config file")
config = None
try:
with open("config.json", 'r') as f:
config = json.loads(f.read())
except:
print("Unable to open config file")
exit()
print("Loaded config file\n", json.dumps(config))
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
connHandler = NewConnHandler()
# For averaging framerate
t0 = time.time()
t1 = 0
last10 = [0] * 10
it = 0
frames = 0
# adjust this to point to your downloaded/trained model
model_path = '/home/nvidia/Documents/SalamiSense/snapshots_1/resnet50_csv_inference.h5'
#model_path = '/home/nvidia/Documents/SalamiSense/resnet50_csv_inference10.h5'
print("loading CNN model")
# load retinanet model
bThread = BoxThread(model_path)
bThread.start()
theta, meanStd = Reg.loadModel("noisy_sensor_model.csv")
gui = initGUI()
thresh = 0
pdist = 0
while keyPress != ord('q'):
connHandler.tick()
# New client connection established
if connHandler.getState() == connHandler.STATE_FINISHED:
newClient = connHandler.getClient()
if newClient is not None:
cids = [citer.getID() for citer in clients]
c = NodeData()
c.setClient(newClient)
for i in range(1000):
if not i in cids:
c.setID(i)
break
clients.append(c)
nindex = len(clients) - 1
bThread.addClient(c)
statusStr, statusColor = connHandler.getStatusStr()
gui.setButtonText("newnode", statusStr, statusColor)
gui.update()
gui.setFrameText("nodeno", "%d/%d" % (nindex + 1, len(clients)))
if not clients:
gui.setFrameImage("vid", [])
gui.setFrameImage("depth", [])
continue
# Receive color and depth data from each client
for c in clients:
data = serv.receiveData(c.getClient(), Roni.TYPE_RGB)
depth = serv.receiveData(c.getClient(), Roni.TYPE_DEPTH)
sense = serv.receiveData(c.getClient(), Roni.TYPE_EDGE)
vfrac = 0
vertStr = None
for i in range(8):
vertStr = serv.receiveData(c.getClient(), Roni.TYPE_3D_0 + i)
if vertStr is not None and vertStr:
vertStr = pickle.loads(vertStr)
vfrac = i
break
# Decode data to images and sensor readings
if data is not None and data:
img = Coppa.decodeColorFrame(data)
c.setColor(img.copy())
if depth is not None and depth:
depImg = Coppa.decodeDepthFrame(depth)
c.setDepth(depImg.copy())
if sense is not None and sense:
sense = pickle.loads(sense)
if len(sense) > 3:
c.setCO2(sense[0])
c.setTVOC(sense[1])
c.setTemp(sense[2])
c.setHumidity(sense[3])
if vertStr is not None and len(vertStr) > 0:
c.setVertex(vertStr, vfrac)
# Skip first 10 frames, just in case
frames += 1
if frames < 10: continue
newThresh = gui.getEntryValue("thresh")
try:
thresh = int(newThresh)
except:
pass
newPDist = gui.getEntryValue("pdist")
try:
pdist = float(newPDist)
except:
pass
newTrgtTemp = gui.getEntryValue("temp")
try:
config["Sensors"]["TargetOccupTemp"] = int(newTrgtTemp)
except:
pass
# get copies of color and depth frames
img = clients[nindex].getColor().copy()
depImg = clients[nindex].getDepth().copy()
if len(img) < 1 or len(depImg) < 1:
continue
fVert = clients[nindex].getVertex()
# Display calibration points
refXY = clients[nindex].getRefXY()
for pt in refXY:
cv2.circle(img, (pt[0], pt[1]), 4, (0, 255, 255), 1)
img[pt[1]][pt[0]] = [0, 255, 255]
# Find valid boxes and add to or update people list
people = []
for c in clients:
boxes, scores, labels = c.getResults()
if len(boxes) > 0:
for box, score, label in zip(boxes[0], scores[0], labels[0]):
# scores are sorted so we can break at first low score found
if score < (thresh / 100):
break
Peeps.AddToPeople(c, box, people, pdist, score)
totalPeeps = len(people)
frameBoxes, scores, isDupe = Peeps.GetIDBoxes(nindex, people)
framePeeps = len(frameBoxes)
# Visualize People found in frame
for i in range(framePeeps):
# color based on if appears in multiple frames
color = (0, 0, 255)
if isDupe[i]: color = (0, 255, 0)
b = frameBoxes[i].astype(int)
draw_box(img, b, color=color)
# draw center point of box
x = int((frameBoxes[i][0] + frameBoxes[i][2]) / 2)
y = int((frameBoxes[i][1] + frameBoxes[i][3]) / 2)
cv2.circle(img, (x, y), 3, (255, 255, 0), -1)
if showConf:
draw_caption(img, b, "%0.0f%%" % (float(scores[i]) * 100))
# Calculate frame rate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
t0 = t1
cv2.putText(img, "stream FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
# Sensor readings
fCO2 = clients[nindex].getCO2()
fTVOC = clients[nindex].getTVOC()
fTemp = clients[nindex].getTemp()
fHum = clients[nindex].getHumidity()
# Regression
# [people, co2, tvoc, tempr, humid]
regInput = [totalPeeps, fCO2, fTVOC, 70, fHum]
res = Reg.runThroughModel(regInput, theta, meanStd)
# Do HVAC stuff
ctrl = controlHVAC(config, res, 70) + ("(%.2f)" % res)
gui.setFrameText("HVAC", ctrl)
# Display sensor readings on GUI
gui.setFrameText("people",
"%d/%d (frame/total)" % (framePeeps, totalPeeps))
gui.setFrameText("temp", "%.2f ºF" % fTemp)
gui.setFrameText("humid", "%.2f%%" % fHum)
gui.setFrameText("co2", "%.0f ppm" % fCO2)
gui.setFrameText("tvoc", "%.0f ppb" % fTVOC)
# Display color and depth images (depth picture-in-picture)
imRGB = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
depRGB = cv2.cvtColor(depImg, cv2.COLOR_BGR2RGB)
guiImg = ImageTk.PhotoImage(Image.fromarray(imRGB))
depImg = ImageTk.PhotoImage(
Image.fromarray(cv2.resize(depRGB, (160, 120))))
gui.setFrameImage("vid", guiImg)
gui.setFrameImage("depth", depImg)
# Close server and display window
bThread.stop()
bThread.join()
serv.close()
cv2.destroyAllWindows()
def main():
global keyPress
global gui
global showConf
global clients
global nindex
global connHandler
global serv
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
connHandler = NewConnHandler()
#connHandler.newClient(serv)
#while(connHandler.getState() != connHandler.STATE_FINISHED):
# connHandler.tick()
#cl = connHandler.getClient()
t0 = time.time()
t1 = 0
last10 = [0] * 10
it = 0
frames = 0
# adjust this to point to your downloaded/trained model
model_path = '/home/nvidia/Documents/SalamiSense/snapshots_1/resnet50_csv_inference.h5'
#model_path = '/home/nvidia/Documents/SalamiSense/resnet50_csv_inference10.h5'
print("loading model")
# load retinanet model
bThread = BoxThread(model_path)
bThread.start()
# load label to names mapping for visualization purposes
labels_to_names = {0: 'person'}
colorizer = rs.colorizer()
gui = initGUI()
thresh = 0
while keyPress != ord('q'):
connHandler.tick()
if connHandler.getState() == connHandler.STATE_FINISHED:
clients.append(connHandler.getClient())
nindex += 1
gui.setButtonText("newnode", connHandler.getStatusStr())
gui.update()
if not clients:
continue
data = []
depth = []
while not data or not depth:
if not data:
data = serv.receiveData(clients[nindex], Roni.TYPE_RGB)
if not depth:
depth = serv.receiveData(clients[nindex], Roni.TYPE_DEPTH)
frames += 1
if frames < 10: continue
# get image
img = Coppa.decodeColorFrame(data)
depImg = Coppa.decodeDepthFrame(depth)
if depImg is None: continue
newThresh = gui.getEntryValue("thresh")
try:
thresh = int(newThresh)
except:
pass
bThread.setData(img)
boxes, scores, labels = bThread.getBoxes()
bc = 0
if len(boxes) > 0:
for box, score, label in zip(boxes[0], scores[0], labels[0]):
# scores are sorted so we can break
if score < (thresh / 100):
break
bc += 1
color = label_color(label)
b = box.astype(int)
draw_box(img, b, color=color)
if showConf:
draw_caption(img, b, "%s" % score)
# Calculate frame rate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
t0 = t1
cv2.putText(img, "stream FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
gui.setFrameText("people", "%d" % bc)
imRGB = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
depRGB = cv2.cvtColor(depImg, cv2.COLOR_BGR2RGB)
guiImg = ImageTk.PhotoImage(Image.fromarray(imRGB))
depImg = ImageTk.PhotoImage(
Image.fromarray(cv2.resize(depRGB, (160, 120))))
gui.setFrameImage("vid", guiImg)
gui.setFrameImage("depth", depImg)
# Display
# Close server and display window
bThread.stop()
bThread.join()
serv.close()
cv2.destroyAllWindows()
def main():
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
cl = serv.getClient()
cv2.namedWindow('RoniStream', cv2.WINDOW_AUTOSIZE)
t0 = time.time()
t1 = 0
last10 = [0] * 10
it = 0
frames = 0
framerates = []
keyPressed = None
while keyPressed != ord('q'):
data = serv.receiveData(cl, Roni.TYPE_RGB)
if len(data) > 0:
frames += 1
# get image, copy, and edge detect
img = Coppa.decodeColorFrame(data)
orig = np.copy(img)
w = img.shape[1]
h = img.shape[0]
#edge detect
edge = cv2.Canny(img, 100, 400)
for i, row in enumerate(edge):
for j, px in enumerate(row):
if px:
img[i][j] = [0, 255, 64]
# Calculate framerate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
if 550 > frames >= 50: framerates.append(fps)
t0 = t1
cv2.putText(img, "FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
# Combine original, edge dectect, and overlay images
edge3 = np.empty_like(img)
edge3 = cv2.cvtColor(edge, cv2.COLOR_GRAY2BGR)
img = cv2.copyMakeBorder(img, 0, 0, int(w / 2), int(w / 2),
cv2.BORDER_CONSTANT, (0, 0, 0))
combined = np.hstack((orig, edge3))
combined = np.vstack((combined, img))
# Display
cv2.imshow('RoniStream', combined)
keyPressed = cv2.waitKey(1)
#with open("last1000frames_cpu.csv", 'w') as fi:
# for n in framerates:
# fi.write("%f," % n)
print(np.average(framerates))
# Close server and display window
serv.close()
cv2.destroyAllWindows()
def main():
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
cl = serv.getClient()
cv2.namedWindow('RoniStream', cv2.WINDOW_AUTOSIZE)
t0 = time.time()
t1 = 0
last10 = [0]*10
it = 0
frames = 0
keyPress = None
# adjust this to point to your downloaded/trained model
model_path = '/home/nvidia/Documents/SalamiSense/snapshots_1/resnet50_csv_inference.h5'
#model_path = '/home/nvidia/Documents/SalamiSense/resnet50_csv_inference10.h5'
print("loading model")
# load retinanet model
bThread = BoxThread(model_path)
bThread.start()
# load label to names mapping for visualization purposes
labels_to_names = {0: 'person'}
while keyPress != ord('q'):
data = []
depth = []
while not data or not depth:
if not data:
data = serv.receiveData(cl, Roni.TYPE_RGB)
if not depth:
depth = serv.receiveData(cl, Roni.TYPE_DEPTH)
frames += 1
if frames < 10: continue
# get image
img = Coppa.decodeColorFrame(data)
depImg = Coppa.decodeDepthFrame(depth)
if depImg is None: continue
depImg = np.asanyarray(depImg)
depColorMap = cv2.applyColorMap(cv2.convertScaleAbs(depImg, alpha=0.03), cv2.COLORMAP_JET)
bThread.setData(img)
boxes, scores, labels = bThread.getBoxes()
bc = 0
if len(boxes) > 0:
for box, score, label in zip(boxes[0], scores[0], labels[0]):
# scores are sorted so we can break
if score < 0.6:
break
bc += 1
color = label_color(label)
b = box.astype(int)
draw_box(img, b, color=color)
draw_caption(img, b, "%s" % score)
# Calculate frame rate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
t0 = t1
cv2.putText(img, "stream FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
cv2.putText(img, "People: %d" % bc, (0, 60),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
combined = np.hstack((img, depColorMap))
# Display
cv2.imshow('RoniStream', combined)
keyPress = cv2.waitKey(1)
# Close server and display window
bThread.stop()
bThread.join()
serv.close()
cv2.destroyAllWindows()
def main():
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
cl = serv.getClient()
cv2.namedWindow('RoniStream', cv2.WINDOW_AUTOSIZE)
t0 = time.time()
t1 = 0
last10 = [0] * 10
it = 0
frames = 0
keyPress = None
# set the modified tf session as backend in keras
keras.backend.tensorflow_backend.set_session(get_session())
# adjust this to point to your downloaded/trained model
model_path = '/home/nvidia/Documents/SalamiSense/snapshots_1/resnet50_csv_inference.h5'
#model_path = '/home/nvidia/Documents/SalamiSense/resnet50_csv_inference.h5'
print("loading model")
# load retinanet model
model = models.load_model(model_path, backbone_name='resnet50')
# load label to names mapping for visualization purposes
labels_to_names = {0: 'person'}
while keyPress != ord('q'):
data = serv.receiveData(cl, Roni.TYPE_RGB)
if len(data) > 0:
frames += 1
if frames < 50: continue
# get image
img = Coppa.decodeColorFrame(data)
img_copy = img.copy()
img = preprocess_image(img)
boxes, scores, labels = model.predict_on_batch(
np.expand_dims(img, axis=0))
for box, score, label in zip(boxes[0], scores[0], labels[0]):
# scores are sorted so we can break
if score < 0.6:
break
color = label_color(label)
b = box.astype(int)
draw_box(img_copy, b, color=color)
draw_caption(img_copy, b, "%s" % score)
# Calculate frame rate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
t0 = t1
cv2.putText(img_copy, "FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
# Display
cv2.imshow('RoniStream', img_copy)
keyPress = cv2.waitKey(1)
# Close server and display window
serv.close()
cv2.destroyAllWindows()
def main():
print("Initializing vid server")
serv = Roni.RoniServer()
serv.listen()
cl = serv.getClient()
cv2.namedWindow('RoniStream', cv2.WINDOW_AUTOSIZE)
mod = GC.makeGpuFun("func", funcStr)
t0 = time.time()
t1 = 0
last10 = [0] * 10
it = 0
framerates = []
frames = 0
keyPress = None
disco = 0
while keyPress != ord('q'):
data = serv.receiveData(cl, Roni.TYPE_RGB)
if keyPress == ord('d'):
disco = 0 if disco else 1
if len(data) > 0:
frames += 1
# get image, copy, and edge detect
img = Coppa.decodeColorFrame(data)
orig = np.copy(img)
edge = cv2.Canny(img, 100, 400)
w = img.shape[1]
h = img.shape[0]
RGB = RB.tick(15)
# Copy data to GPU
imgGPU = GC.gpuMemcpy(img, img.nbytes)
edgeGPU = GC.gpuMemcpy(edge, edge.nbytes)
# Launch kernel
grid = (int((w + 15) / 16), int((h + 15) / 16), 1)
GC.execute(mod,
imgGPU,
edgeGPU,
np.int32(h),
np.int32(w),
np.int32(RGB[2]),
np.int32(RGB[1]),
np.int32(RGB[0]),
np.int32(disco),
grid=grid)
# Copy memory back from GPU
GC.gpuMemget(img, imgGPU)
# Calculate frame rate and display on image
t1 = time.time()
last10[it] = 1.0 / (t1 - t0)
it = (it + 1) % 10
fps = np.average(last10)
if 550 > frames >= 50: framerates.append(fps)
t0 = t1
cv2.putText(img, "FPS: %.2f" % fps, (0, 30),
cv2.FONT_HERSHEY_PLAIN, 1.2, (0, 0, 255))
# Combine original, edge detect, and overlay images
edge3 = np.empty_like(img)
edge3 = cv2.cvtColor(edge, cv2.COLOR_GRAY2BGR)
img = cv2.copyMakeBorder(img, 0, 0, int(w / 2), int(w / 2),
cv2.BORDER_CONSTANT, (0, 0, 0))
combined = np.hstack((orig, edge3))
combined = np.vstack((combined, img))
# Display
cv2.imshow('RoniStream', combined)
keyPress = cv2.waitKey(1)
#with open("last1000frames.csv", 'w') as fi:
# for n in framerates:
# fi.write("%f," % n)
print(np.average(framerates))
# Close server and display window
serv.close()
cv2.destroyAllWindows()