180 lines
7.7 KiB
Python
180 lines
7.7 KiB
Python
"""
|
|
Many thanks to https://github.com/vardanagarwal/Proctoring-AI/blob/master/coco models/tflite mobnetv1 ssd
|
|
"""
|
|
|
|
|
|
import cv2
|
|
import numpy as np
|
|
import multiprocessing
|
|
import threading
|
|
import os
|
|
from copy import copy
|
|
from time import sleep
|
|
|
|
from . import Algorithm
|
|
|
|
""" Check if neural network accelerator is existing """
|
|
try_edgetpu = True
|
|
try:
|
|
if not try_edgetpu:
|
|
raise Exception()
|
|
from pycoral.adapters import common
|
|
from pycoral.adapters import detect
|
|
from pycoral.utils.dataset import read_label_file
|
|
from pycoral.utils.edgetpu import make_interpreter, list_edge_tpus
|
|
if len(list_edge_tpus()) == 0:
|
|
raise Exception()
|
|
engine = "edgetpu"
|
|
except Exception as e:
|
|
import tensorflow as tf
|
|
engine = "tflite"
|
|
|
|
|
|
class Detector:
|
|
def __init__(self):
|
|
self.category_index = self.create_category_index()
|
|
if engine == "tflite":
|
|
self.num_threads = int(multiprocessing.cpu_count())
|
|
print("Self using %s threads for object detection" % self.num_threads)
|
|
self.interpreter = tf.lite.Interpreter(
|
|
model_path="data" + os.sep + "ssd_mobilenet_v2_coco_quant_postprocess.tflite", num_threads=self.num_threads
|
|
)
|
|
self.interpreter.allocate_tensors()
|
|
# Get input and output tensors.
|
|
self.input_details = self.interpreter.get_input_details()
|
|
self.output_details = self.interpreter.get_output_details()
|
|
elif engine == "edgetpu":
|
|
print("Running with edge tpu")
|
|
self.interpreter = make_interpreter("data" + os.sep + "ssd_mobilenet_v2_coco_quant_postprocess_edgetpu.tflite")
|
|
self.interpreter.allocate_tensors()
|
|
# Get input and output tensors.
|
|
self.input_details = self.interpreter.get_input_details()
|
|
self.output_details = self.interpreter.get_output_details()
|
|
|
|
def create_category_index(self, label_path='data' + os.sep + 'labelmap.txt'):
|
|
f = open(label_path)
|
|
category_index = {}
|
|
for i, val in enumerate(f):
|
|
if i != 0:
|
|
val = val[:-1]
|
|
category_index.update({(i - 1): {'id': (i - 1), 'name': val}})
|
|
f.close()
|
|
return category_index
|
|
|
|
def get_output_dict(self, nms=True, iou_thresh=0.5, score_thresh=0.5):
|
|
output_dict = {
|
|
'detection_boxes': self.interpreter.get_tensor(self.output_details[0]['index'])[0],
|
|
'detection_classes': self.interpreter.get_tensor(self.output_details[1]['index'])[0],
|
|
'detection_scores': self.interpreter.get_tensor(self.output_details[2]['index'])[0],
|
|
'num_detections': self.interpreter.get_tensor(self.output_details[3]['index'])[0]
|
|
}
|
|
output_dict['detection_classes'] = output_dict['detection_classes'].astype(np.int64)
|
|
output_dict["detection_classes_name"] = [self.category_index[x] for x in output_dict["detection_classes"]]
|
|
if nms and engine == "tflite":
|
|
output_dict = self.apply_nms(output_dict, iou_thresh, score_thresh)
|
|
if nms and engine == "edgetpu":
|
|
valid = np.where(output_dict["detection_scores"] >= score_thresh)[0]
|
|
if valid.size == 0:
|
|
output_dict = {}
|
|
elif valid.size == 1:
|
|
output_dict = {
|
|
'detection_boxes': output_dict["detection_boxes"][valid[0]:valid[0] + 1],
|
|
'detection_classes': output_dict["detection_classes"][valid[0]:valid[0] + 1],
|
|
'detection_scores': output_dict["detection_scores"][valid[0]:valid[0] + 1],
|
|
'detection_classes_name': output_dict["detection_classes_name"][valid[0]:valid[0] + 1],
|
|
'num_detections': 1,
|
|
}
|
|
else:
|
|
output_dict = {
|
|
'detection_boxes': output_dict["detection_boxes"][valid],
|
|
'detection_classes': output_dict["detection_classes"][valid],
|
|
'detection_scores': output_dict["detection_scores"][valid],
|
|
'detection_classes_name': [x for i,x in enumerate(output_dict["detection_classes_name"]) if i in valid],
|
|
'num_detections': valid.size,
|
|
}
|
|
return output_dict
|
|
|
|
def apply_nms(self, output_dict, iou_thresh=0.5, score_thresh=0.5):
|
|
q = 90 # no of classes
|
|
num = int(output_dict['num_detections'])
|
|
boxes = np.zeros([1, num, q, 4])
|
|
scores = np.zeros([1, num, q])
|
|
# val = [0]*q
|
|
for i in range(num):
|
|
# indices = np.where(classes == output_dict['detection_classes'][i])[0][0]
|
|
boxes[0, i, output_dict['detection_classes'][i], :] = output_dict['detection_boxes'][i]
|
|
scores[0, i, output_dict['detection_classes'][i]] = output_dict['detection_scores'][i]
|
|
nmsd = tf.image.combined_non_max_suppression(
|
|
boxes=boxes, scores=scores, max_output_size_per_class=num, max_total_size=num, iou_threshold=iou_thresh,
|
|
score_threshold=score_thresh, pad_per_class=False, clip_boxes=False
|
|
)
|
|
valid = nmsd.valid_detections[0].numpy()
|
|
output_dict = {
|
|
'detection_boxes': nmsd.nmsed_boxes[0].numpy()[:valid],
|
|
'detection_classes': nmsd.nmsed_classes[0].numpy().astype(np.int64)[:valid],
|
|
'detection_scores': nmsd.nmsed_scores[0].numpy()[:valid],
|
|
'detection_classes_name': output_dict["detection_classes_name"][:valid]
|
|
}
|
|
return output_dict
|
|
|
|
def make_inference(self, img, nms=True, score_thresh=0.5, iou_thresh=0.5):
|
|
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
|
img_rgb = cv2.resize(img_rgb, (300, 300), cv2.INTER_AREA)
|
|
img_rgb = img_rgb.reshape([1, 300, 300, 3])
|
|
self.interpreter.set_tensor(self.input_details[0]['index'], img_rgb)
|
|
self.interpreter.invoke()
|
|
output_dict = self.get_output_dict(nms, iou_thresh, score_thresh)
|
|
return output_dict
|
|
|
|
|
|
class ObjectDetector(Algorithm):
|
|
""" Detects objects """
|
|
|
|
def __init__(self):
|
|
""" Init some values and set seed point to None """
|
|
self.objects = dict()
|
|
self.detection_image = None
|
|
self.lock = threading.Lock()
|
|
self.thread = threading.Thread(target=self._detect, args=[], daemon=True)
|
|
self.thread.start()
|
|
|
|
def process(self, img):
|
|
"""
|
|
Tries to segment a region around the seed point and calculates a new seed point by finding the segments center
|
|
"""
|
|
with self.lock:
|
|
if self.detection_image is None:
|
|
self.detection_image = np.copy(img)
|
|
with self.lock:
|
|
objects = copy(self.objects)
|
|
h, w, c = img.shape
|
|
if "detection_classes_name" in objects.keys():
|
|
for i, cls in enumerate(objects["detection_classes_name"]):
|
|
box = objects["detection_boxes"][i]
|
|
score = objects["detection_scores"][i]
|
|
name = cls["name"]
|
|
y1, x1, y2, x2 = round(box[0] * h), round(box[1] * w), round(box[2] * h), round(box[3] * w)
|
|
img = cv2.rectangle(img, (x1, y1), (x2, y2), color=(0, 0, 0), thickness=2)
|
|
img = cv2.putText(img, "%s: %.2f" % (name, score), (x1, y1), fontFace=cv2.FONT_HERSHEY_SIMPLEX, fontScale=1, color=(0, 0, 0))
|
|
|
|
return img
|
|
|
|
def _detect(self):
|
|
detector = Detector()
|
|
while True:
|
|
with self.lock:
|
|
img = self.detection_image
|
|
if img is None:
|
|
sleep(.033)
|
|
continue
|
|
objects = detector.make_inference(img)
|
|
with self.lock:
|
|
self.objects = objects
|
|
self.detection_image = None
|
|
|
|
def mouse_callback(self, event, x, y, flags, param):
|
|
""" Selects a new seed point"""
|
|
if event == cv2.EVENT_LBUTTONUP:
|
|
pass
|
|
|