畢竟nano pi M4計算能力有限，我們這里先選擇專門為速度優化過最快的網絡SSD，以及經典的faster-rcnn作對比，再加上能顯示mask的高端網絡，，，
事實上yolo v3剛出來，比SSD更快，而faster rcnn相對來說運行慢的多了，后面可以都嘗試對比一下，目前先把基線系統搭建好。
環境：
系統：rk3399-sd-friendlydesktop-bionic-4.4-arm64
python2.7
tensorflow1.8.0
opencv
平台：
nano pi M4 2G版
logi攝像頭
TF卡32G
1、進入這個網站下載適合的tensorflow版本：https://github.com/lhelontra/tensorflow-on-arm/releases/download/v1.8.0這里使用tensorflow1.8.0；
sudo pip install tensorflow-1.8.0-cp27-none-linux_aarch64.whl
或者：在線下載安裝wget https://github.com/lhelontra/tensorflow-on-arm/releases/download/v1.8.0/tensorflow-1.8.0-cp27-none-linux_aarch64.whl
sudo pip install tensorflow-1.8.0-cp27-none-linux_aarch64.whl
2、安裝matplotlib庫
sudo pip install matplotlib
3、opencv安裝sudo pip install opencv-python
4、下載tensorflow提供的models API並解壓，下載模型git clone https://github.com/tensorflow/models.git
5、下載用COCO訓練集預訓練的模型
下載訓練好的模型並放到上一步models_master下的object_detection/models目錄
6、Protobuf 安裝與配置
下載地址： https://github.com/google/protobuf/releases
我們這里下載最新版本
protobuf-all-3.5.1.tar.gz
tar -xf protobuf-all-3.5.1.tar.gz cd protobuf-3.5.1 ./configure make make check ->這一步是檢查編譯是否正確，耗時非常長，可略過 sudo make install sudo ldconfig ->更新庫搜索路徑，否則可能找不到庫文件
將proto格式的數據轉換為python格式，從而可以在python腳本中調用，進入目錄models-master/research，運行：
protoc object_detection/protos/*.proto --python_out=.
轉換完畢后可以看到在object_detection/protos/目錄下多了許多*.py文件。
7、測試代碼
import numpy as np import os import sys import tarfile import tensorflow as tf import cv2 import time from collections import defaultdict # This is needed since the notebook is stored in the object_detection folder. sys.path.append("../..") from object_detection.utils import label_map_util from object_detection.utils import visualization_utils as vis_util # What model to download. MODEL_NAME = 'ssd_mobilenet_v1_coco_2018_01_28' # Path to frozen detection graph. This is the actual model that is used for the object detection. PATH_TO_CKPT = MODEL_NAME + '/frozen_inference_graph.pb' # List of the strings that is used to add correct label for each box. PATH_TO_LABELS = os.path.join('/home/pi/models/research/object_detection/data', 'mscoco_label_map.pbtxt') model_path = "/home/pi/models/research/object_detection/models/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt" #extract the ssd_mobilenet start = time.clock() NUM_CLASSES = 90 end= time.clock() print('load the model' ,(end -start)) detection_graph = tf.Graph() with detection_graph.as_default(): od_graph_def = tf.GraphDef() with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid: serialized_graph = fid.read() od_graph_def.ParseFromString(serialized_graph) tf.import_graph_def(od_graph_def, name='') label_map = label_map_util.load_labelmap(PATH_TO_LABELS) categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True) category_index = label_map_util.create_category_index(categories) cap = cv2.VideoCapture(0) with detection_graph.as_default(): with tf.Session(graph=detection_graph) as sess: writer = tf.summary.FileWriter("logs/", sess.graph) sess.run(tf.global_variables_initializer()) loader = tf.train.import_meta_graph(model_path + '.meta') loader.restore(sess, model_path) while(1): start = time.clock() ret, frame = cap.read() if cv2.waitKey(1) & 0xFF == ord('q'): break image_np =frame # the array based representation of the image will be used later in order to prepare the # result image with boxes and labels on it. # Expand dimensions since the model expects images to have shape: [1, None, None, 3] image_np_expanded = np.expand_dims(image_np, axis=0) image_tensor = detection_graph.get_tensor_by_name('image_tensor:0') # Each box represents a part of the image where a particular object was detected. boxes = detection_graph.get_tensor_by_name('detection_boxes:0') # Each score represent how level of confidence for each of the objects. # Score is shown on the result image, together with the class label. scores = detection_graph.get_tensor_by_name('detection_scores:0') classes = detection_graph.get_tensor_by_name('detection_classes:0') num_detections = detection_graph.get_tensor_by_name('num_detections:0') # Actual detection. (boxes, scores, classes, num_detections) = sess.run( [boxes, scores, classes, num_detections], feed_dict={image_tensor: image_np_expanded}) # Visualization of the results of a detection. vis_util.visualize_boxes_and_labels_on_image_array( image_np, np.squeeze(boxes), np.squeeze(classes).astype(np.int32), np.squeeze(scores), category_index, use_normalized_coordinates=True, line_thickness=6) end = time.clock() print 'One frame detect take time:' ,end - start cv2.imshow("capture", image_np) print('after cv2 show') cv2.waitKey(1) cap.release() cv2.destroyAllWindows()
保存為 test.py，到目錄models-master/research/object_detection/models下。
8、運行：
進入models-master/research/object_detection/models運行命令：
sudo chmod 666 /dev/video0
python detect.py
9、測試結果：可以看到SSD模型加載花了3秒左右，識別一張圖在2秒左右。
安裝環境是會出現pip安裝超超時等等問題，解決辦法：sudo pip install --index-url https://pypi.douban.com/simple matplotlib
sudo apt-get install protobuf-compiler
安裝opencv時 sudo apt-get upgrade sudo apt-get install python2.7-dev sudo apt-get install python-opencv