async def main():
# Set minimum confidence level required to draw bounding box; default is 60%
minimum_confidence = 0.60
# Set image folder (folder will be mirrored on the cloud); should be determined before running; default is 0
image_folder_number = 0
path = f"./cloudcam_session{image_folder_number}"
while os.path.isdir(path) or os.path.isfile(path):
print("Incrementing folder number")
image_folder_number += 1
path = f"./cloudcam_session{image_folder_number}"
print(image_folder_number)
os.mkdir(f"./cloudcam_session{image_folder_number}")
cloud_manager = CloudManager()
await cloud_manager.connect()
# Read label file and load labels into a variable
with open("./Sample_TFLite_model/labelmap.txt") as file:
labels = [line.strip() for line in file.readlines()]
if labels[0] == '???':
del (labels[0])
print("Loading model into interpreter...")
interpreter = Interpreter(model_path="./Sample_TFLite_model/detect.tflite")
# Getting useful information on the model for later use
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
model_input_height = input_details[0]["shape"][1]
model_input_width = input_details[0]["shape"][2]
# Preparing model inside interpreter
interpreter.allocate_tensors()
# Getting the video stream
video_handler = VideoStream()
frame_counter = 0
for frame in video_handler.capture_stream():
image = frame.array
print("Performing image pre-processing...")
# image_path = "./<YOUR TEST IMAGE>.jpg" # Can change accordingly
# image = cv2.imread( image_path )
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_rows, image_columns, image_color_channels = image.shape # Get original image height / width information
image_resized = cv2.resize(
image_rgb, (model_input_width, model_input_height
)) # Resize image to fit model's required input shape
model_input_data = np.expand_dims(image_resized, axis=0)
print("Performing inference...")
interpreter.set_tensor(input_details[0]["index"], model_input_data)
interpreter.invoke() # Executes the inference
boxes = interpreter.get_tensor(output_details[0]['index'])[0]
classes = interpreter.get_tensor(output_details[1]["index"])[0]
scores = interpreter.get_tensor(output_details[2]["index"])[0]
for i in range(len(scores)):
object_detected = labels[int(classes[i])]
confidence = scores[i]
if confidence > minimum_confidence and object_detected == "person": # can omit section after "and" to detect multiple objects
frame_counter += 1
print("Object:", object_detected)
print("Confidence:", scores[i])
print("")
# Get x and y coordinates of bounding box, correcting for inference results that lay off the original image
y_lower = int(max(1, (boxes[i][0] * image_rows)))
y_upper = int(max(image_rows, (boxes[i][2] * image_rows)))
x_lower = int(max(1, (boxes[i][1] * image_columns)))
x_upper = int(max(image_columns,
(boxes[i][3] * image_columns)))
# Draw bounding box on the image
cv2.rectangle(image, (x_lower, y_lower), (x_upper, y_upper),
(255, 0, 0), 3)
# cv2.imshow("Inference Result", image)
new_image_path = f"cloudcam_session{image_folder_number}/cloudcam_image{frame_counter}.jpg"
print("Saving image locally to Pi...")
cv2.imwrite(new_image_path,
image) # Can disable if Pi's memory is constrained
confidence_string = str(round(confidence * 100, 1)) + '%'
# Create and send message to cloud
D2C_message = f"Detected {object_detected}"
await cloud_manager.send_message(confidence_string,
D2C_message)
await cloud_manager.store_in_blob(new_image_path,
confidence_string)
# Clear stream for next frame
video_handler.array_capture.truncate(0)
time.sleep(2)
