You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 

197 lines
8.3 KiB

#!/usr/bin/env python3
import cv2
import numpy as np
from scipy.spatial.transform import Rotation
import deep_models_shared_python3 as dm
class ObjectDetector:
def __init__(self, models_directory, use_tiny_yolo3 = True, confidence_threshold=0.2, use_neural_compute_stick=False):
# Load the models
models_dir = models_directory + 'darknet/'
print('Using the following directory to load object detector models:', models_dir)
if use_tiny_yolo3:
model_filename = models_dir + 'tiny_yolo_v3/yolov3-tiny.weights'
config_filename = models_dir + 'tiny_yolo_v3/yolov3-tiny.cfg'
classes_filename = models_dir + 'tiny_yolo_v3/object_detection_classes_yolov3.txt'
input_width = 416
input_height = 416
else:
model_filename = models_dir + 'yolo_v3/yolov3.weights'
config_filename = models_dir + 'yolo_v3/yolov3.cfg'
classes_filename = models_dir + 'yolo_v3/object_detection_classes_yolov3.txt'
input_width = 608
input_height = 608
self.input_width = input_width
self.input_height = input_height
self.confidence_threshold = confidence_threshold
self.non_maximal_suppression = 0.01
self.scale = 0.00392
self.rgb = True
self.mean = (0.0, 0.0, 0.0)
if use_tiny_yolo3:
print('using YOLO V3 Tiny')
else:
print('using YOLO V3')
print('models_dir =', models_dir)
print('model_filename =', model_filename)
print('config_filename =', config_filename)
print('classes_filename =', classes_filename)
classes_file = open(classes_filename, 'rt')
raw_classes_text = classes_file.read()
classes_file.close()
self.object_class_labels = raw_classes_text.rstrip('\n').split('\n')
self.num_object_classes = len(self.object_class_labels)
self.object_detection_model = cv2.dnn.readNet(model_filename, config_filename, 'darknet')
# attempt to use Neural Compute Stick 2
if use_neural_compute_stick:
print('ObjectDetector.__init__: Attempting to use an Intel Neural Compute Stick 2 using the following command: self.object_detection_model.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD)')
self.object_detection_model.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD)
dm.print_model_info(self.object_detection_model, 'object_detection_model')
self.output_layer_names = self.object_detection_model.getUnconnectedOutLayersNames()
def get_landmark_names(self):
return None
def get_landmark_colors(self):
return None
def get_landmark_color_dict(self):
return None
def apply_to_image(self, rgb_image, draw_output=False):
original_height, original_width, num_color = rgb_image.shape
object_image_blob = cv2.dnn.blobFromImage(rgb_image,
1.0,
size=(self.input_width, self.input_height),
swapRB=self.rgb,
ddepth=cv2.CV_8U)
self.object_detection_model.setInput(object_image_blob, scalefactor=self.scale, mean=self.mean)
object_detections = self.object_detection_model.forward(self.output_layer_names)
# object_detections is a list
# YOLO v3 Tiny
# object_detections = [ array with shape (507, 85),
# array with shape (2028, 85) ]
# YOLO v3
# object_detections = [ array with shape (1083, 85),
# array with shape (4332, 85),
# array with shape (17328, 85) ]
# each element of the list has a constant shape RxC
# Each of the R rows represents a detection
# [0:5] (the first 4 numbers) specify a bounding box
# [box_center_x, box_center_y, box_width, box_height], where
# each element is a scalar between 0.0 and 1.0 that can be
# multiplied by the original input image dimensions to recover
# the bounding box in the original image.
# [5:] (the remaining 81 numbers) represent the confidence
# that a particular class was detected in the bounding box (80
# COCO object classes) plus one class that represents the
# background and hence no detection (most likely - my
# interpretation without really looking closely at it).
def bound_x(x_in):
x_out = max(x_in, 0)
x_out = min(x_out, original_width - 1)
return x_out
def bound_y(y_in):
y_out = max(y_in, 0)
y_out = min(y_out, original_height - 1)
return y_out
results = []
for detections in object_detections:
object_class_confidences = detections[:,5:]
best_object_classes = np.argmax(object_class_confidences, axis=1)
# only consider non-background classes
non_background_selector = best_object_classes < self.num_object_classes
detected_objects = detections[non_background_selector]
best_object_classes = best_object_classes[non_background_selector]
# collect and prepare detected objects
for detection, object_class_id in zip(detected_objects, best_object_classes):
confidence = detection[5:][object_class_id]
if confidence > self.confidence_threshold:
class_label = self.object_class_labels[object_class_id]
box_center_x, box_center_y, box_width, box_height = detection[:4]
x_min = (box_center_x - (box_width / 2.0)) * original_width
y_min = (box_center_y - (box_height / 2.0)) * original_height
x_max = x_min + (box_width * original_width)
y_max = y_min + (box_height * original_height)
x_min = bound_x(int(round(x_min)))
y_min = bound_y(int(round(y_min)))
x_max = bound_x(int(round(x_max)))
y_max = bound_y(int(round(y_max)))
box = (x_min, y_min, x_max, y_max)
print(class_label, ' detected')
results.append({'class_id': object_class_id,
'label': class_label,
'confidence': confidence,
'box': box})
output_image = None
if draw_output:
output_image = rgb_image.copy()
for detection_dict in results:
self.draw_detection(output_image, detection_dict)
return results, output_image
def draw_detection(self, image, detection_dict):
font_scale = 0.75
line_color = [0, 0, 0]
line_width = 1
font = cv2.FONT_HERSHEY_PLAIN
class_label = detection_dict['label']
confidence = detection_dict['confidence']
box = detection_dict['box']
x_min, y_min, x_max, y_max = box
output_string = '{0}, {1:.2f}'.format(class_label, confidence)
color = (0, 0, 255)
rectangle_line_thickness = 2 #1
cv2.rectangle(image, (x_min, y_min), (x_max, y_max), color, rectangle_line_thickness)
# see the following page for a helpful reference
# https://stackoverflow.com/questions/51285616/opencvs-gettextsize-and-puttext-return-wrong-size-and-chop-letters-with-low
label_background_border = 2
(label_width, label_height), baseline = cv2.getTextSize(output_string, font, font_scale, line_width)
label_x_min = x_min
label_y_min = y_min
label_x_max = x_min + (label_width + (2 * label_background_border))
label_y_max = y_min + (label_height + baseline + (2 * label_background_border))
text_x = label_x_min + label_background_border
text_y = (label_y_min + label_height) + label_background_border
cv2.rectangle(image, (label_x_min, label_y_min), (label_x_max, label_y_max), (255, 255, 255), cv2.FILLED)
cv2.putText(image, output_string, (text_x, text_y), font, font_scale, line_color, line_width, cv2.LINE_AA)