|
我们现在已经有了可以识别姿势的模型，接下来要做的只是把这个模型和无线灯光控制关联起来就行了。
在我的这个例子中，我使用Aeotec Z-Stick来发送Z-Wave指令，并配有两个GE Z-Wave的室外开关。USB接口接入到NVIDIA TX2人工智能模块，其实NVIDIA的Jestson Nano也能胜任，尽管Jetson Nano所能提供的分辨率要低于我样例中1280x720的分辨率。当Z-Stick插入到ARM设备后，你首先需要把开关调到Z-Wave模式，可能需要多按几下USB Stick上的按钮和灯的开关。
代码并不复杂，基本上就是训练环境再加上一个额外的步骤。现在，我们导入Keras，然后使用清理过的数据训练模型。
- import cv2import pyopenpose as opfrom imutils import translate, rotate, resizeimport openzwavefrom openzwave.option import ZWaveOptionfrom openzwave.network import ZWaveNetwork# make sure these commands get flushed by doing them first, then loading tensorflow...# tensorflow should take enough time to start for these commands to flushoptions = ZWaveOption('/dev/ttyACM0')options.lock()network = ZWaveNetwork(options)import timeimport numpy as npnp.random.seed(1337)import tensorflow as tf# make sure tensorflow doesn't take up all the gpu memoryconf = tf.ConfigProto()conf.gpu_options.allow_growth=Truesession = tf.Session(config=conf)import keras# Custom Params (refer to include/openpose/flags.hpp for more parameters)params = dict()params["model_folder"] = "../../models/"# built in TX2 video capture sourcevs = cv2.VideoCapture("nvarguscamerasrc ! video/x-raw(memory:NVMM), width=(int)1280, height=(int)720,format=(string)NV12, framerate=(fraction)24/1 ! nvvidconv flip-method=0 ! video/x-raw, format=(string)BGRx ! videoconvert ! video/x-raw, format=(string)BGR ! appsink")tposer = keras.models.load_model('dab-tpose-other.h5')# Starting OpenPoseopWrapper = op.WrapperPython()opWrapper.configure(params)opWrapper.start()datum = op.Datum()np.set_printoptions(precision=4)fps_time = 0DAB = 1TPOSE = 2OTHER = 0LIGHTS = 0bounced = time.time()debounce = 3 # wait 3 seconds before allowing another commandwhile True: ret_val, frame = vs.read() datum.cvInputData = frame opWrapper.emplaceAndPop([datum]) # need to be able to see what's going on image = datum.cvOutputData cv2.putText(image, "FPS: %f" % (1.0 / (time.time() - fps_time)), (10, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2) cv2.imshow("Openpose", image) if datum.poseKeypoints.any(): first_input = datum.poseKeypoints try: first_input[:,:,0] = first_input[:,:,0] / 720 first_input[:,:,1] = first_input[:,:,1] / 1280 first_input = first_input[:,:,1:] first_input = first_input.reshape(len(datum.poseKeypoints), 50) except: continue output = tposer.predict_classes(first_input) for j in output: if j == 1: print("dab detected") if LIGHTS == 0 or (time.time() - bounced) < debounce: continue for node in network.nodes: for val in network.nodes[node].get_switches(): network.nodes[node].set_switch(val, False) LIGHTS = 0 bounced = time.time() elif j == 2: print("tpose detected") if LIGHTS == 1 or (time.time() - bounced) < debounce: continue for node in network.nodes: for val in network.nodes[node].get_switches(): network.nodes[node].set_switch(val, True) LIGHTS = 1 bounced = time.time() fps_time = time.time() # quit with a q keypress, b or m to save data key = cv2.waitKey(1) & 0xFF if key == ord("q"): break# clean up after yourselfvs.release()cv2.destroyAllWindows()
到了这一步，工作基本上就算完成了！
我们成功地训练了一个用于识别嘻哈超人舞步、T-Pose舞步的神经网络模型，然后我们可以让它根据我们的舞步来制造可互动的灯。
太棒了，给自己点个赞！
后记
所有代码、模型以及训练数据都免费公布在Github上。 （编辑：厦门网）
【声明】本站内容均来自网络，其相关言论仅代表作者个人观点，不代表本站立场。若无意侵犯到您的权利，请及时与联系站长删除相关内容!
|
