devbox
AllinToyou
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

树莓派部署YOLOv5模型_树莓派yolov5

作者：AllinToyou | 2024-02-26 21:34:08

赞

踩

树莓派yolov5

本文章是关于树莓派部署YOLOv5s模型，实际测试效果的FPS仅有0.15，不够满足实际检测需要，各位大佬可以参考参考。

1、在树莓派中安装opencv（默认安装好python3）

  1. # 直接安装
  2. # 安装依赖软件
  3. sudo apt-get install -y libopencv-dev python3-opencv
  4. sudo apt-get install libatlas-base-dev
  5. sudo apt-get install libjasper-dev
  6. sudo apt-get install libqtgui4
  7. sudo apt-get install python3-pyqt5
  8. sudo apt install libqt4-test
  9. # 安装Python 包
  10. pip3 install opencv-python

2、导出onnx模型

从YOLOv5官网下载源代码和YOLOv5s.pt文件

YOLOv5官网

YOLOv5s.pt下载

按照作者提示安装环境，使用它自带的export.py将YOLOv5s.pt转为YOLOv5s.onnx，安装好环境后，在终端输入以下命令即可自动生成。

python export.py --weights yolov5s.pt --include onnx

3.测试

        可以先在电脑上测试一下，使用如下代码测试上述转换的模型能否使用，假如成功即可将下述代码和上述生成的YOLOv5s.onnx模型直接移动到树莓派中进行测试。

  1. # 图片检测
  2. import cv2
  3. import numpy as np
  4. import time
  5. def plot_one_box(x, img, color=None, label=None, line_thickness=None):
  6.     """
  7.     description: Plots one bounding box on image img,
  8.                  this function comes from YoLov5 project.
  9.     param: 
  10.         x:      a box likes [x1,y1,x2,y2]
  11.         img:    a opencv image object
  12.         color:  color to draw rectangle, such as (0,255,0)
  13.         label:  str
  14.         line_thickness: int
  15.     return:
  16.         no return
  17.     """
  18.     tl = (
  19.         line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1
  20.     )  # line/font thickness
  21.     color = color or [random.randint(0, 255) for _ in range(3)]
  22.     c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
  23.     cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
  24.     if label:
  25.         tf = max(tl - 1, 1)  # font thickness
  26.         t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
  27.         c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
  28.         cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA)  # filled
  29.         cv2.putText(
  30.             img,
  31.             label,
  32.             (c1[0], c1[1] - 2),
  33.             0,
  34.             tl / 3,
  35.             [225, 255, 255],
  36.             thickness=tf,
  37.             lineType=cv2.LINE_AA,
  38.         )
  39.  
  40. def post_process_opencv(outputs,model_h,model_w,img_h,img_w,thred_nms,thred_cond):
  41.     
  42.     conf = outputs[:,4].tolist()
  43.     c_x = outputs[:,0]/model_w*img_w
  44.     c_y = outputs[:,1]/model_h*img_h
  45.     w  = outputs[:,2]/model_w*img_w
  46.     h  = outputs[:,3]/model_h*img_h
  47.     p_cls = outputs[:,5:]
  48.     if len(p_cls.shape)==1:
  49.         p_cls = np.expand_dims(p_cls,1)
  50.     cls_id = np.argmax(p_cls,axis=1)
  51.  
  52.     p_x1 = np.expand_dims(c_x-w/2,-1)
  53.     p_y1 = np.expand_dims(c_y-h/2,-1)
  54.     p_x2 = np.expand_dims(c_x+w/2,-1)
  55.     p_y2 = np.expand_dims(c_y+h/2,-1)
  56.     areas = np.concatenate((p_x1,p_y1,p_x2,p_y2),axis=-1)
  57.     print(areas.shape) 
  58.     areas = areas.tolist()
  59.     ids = cv2.dnn.NMSBoxes(areas,conf,thred_cond,thred_nms) 
  60.     return  np.array(areas)[ids],np.array(conf)[ids],cls_id[ids]
  61.  
  62.  
  63. def infer_image(net,img0,model_h,model_w,thred_nms=0.4,thred_cond=0.5):
  64.  
  65.     img = img0.copy()
  66.     img = cv2.resize(img,[model_h,model_w])
  67.     blob = cv2.dnn.blobFromImage(img, scalefactor=1/255.0, swapRB=True)
  68.     net.setInput(blob)
  69.     outs = net.forward()[0]
  70.     print(outs[0])
  71.     det_boxes,scores,ids = post_process_opencv(outs,model_h,model_w,img0.shape[0],img0.shape[1],thred_nms,thred_cond)
  72.     return det_boxes,scores,ids
  73.  
  74.  
  75. if __name__=="__main__":
  76.     dic_labels= {0: 'person',
  77.   1: 'bicycle',
  78.   2: 'car',
  79.   4: 'airplane',
  80.   5: 'bus',
  81.   6: 'train',
  82.   7: 'truck',
  83.   8: 'boat',
  84.   9: 'traffic light',
  85.   10: 'fire hydrant',
  86.   11: 'stop sign',
  87.   12: 'parking meter',
  88.   13: 'bench',
  89.   14: 'bird',
  90.   15: 'cat',
  91.   16: 'dog',
  92.   17: 'horse',
  93.   18: 'sheep',
  94.   19: 'cow',
  95.   20: 'elephant',
  96.   21: 'bear',
  97.   22: 'zebra',
  98.   23: 'giraffe',
  99.   24: 'backpack',
  100.   25: 'umbrella',
  101.   26: 'handbag',
  102.   27: 'tie',
  103.   28: 'suitcase',
  104.   29: 'frisbee',
  105.   30: 'skis',
  106.   31: 'snowboard',
  107.   32: 'sports ball',
  108.   33: 'kite',
  109.   34: 'baseball bat',
  110.   35: 'baseball glove',
  111.   36: 'skateboard',
  112.   37: 'surfboard',
  113.   38: 'tennis racket',
  114.   39: 'bottle',
  115.   40: 'wine glass',
  116.   41: 'cup',
  117.   42: 'fork',
  118.   43: 'knife',
  119.   44: 'spoon',
  120.   45: 'bowl',
  121.   46: 'banana',
  122.   47: 'apple',
  123.   48: 'sandwich',
  124.   49: 'orange',
  125.   50: 'broccoli',
  126.   51: 'carrot',
  127.   52: 'hot dog',
  128.   53: 'pizza',
  129.   54: 'donut',
  130.   55: 'cake',
  131.   56: 'chair',
  132.   57: 'couch',
  133.   58: 'potted plant',
  134.   59: 'bed',
  135.   60: 'dining table',
  136.   61: 'toilet',
  137.   62: 'tv',
  138.   63: 'laptop',
  139.   64: 'mouse',
  140.   65: 'remote',
  141.   66: 'keyboard',
  142.   67: 'cell phone',
  143.   68: 'microwave',
  144.   69: 'oven',
  145.   70: 'toaster',
  146.   71: 'sink',
  147.   72: 'refrigerator',
  148.   73: 'book',
  149.   74: 'clock',
  150.   75: 'vase',
  151.   76: 'scissors',
  152.   77: 'teddy bear',
  153.   78: 'hair drier',
  154.   79: 'toothbrush'}
  155.  
  156.     model_h = 640
  157.     model_w = 640
  158.     file_model = 'yolov5s.onnx'
  159.     net = cv2.dnn.readNet(file_model)
  160.     img0 = cv2.imread('3.jpg')
  161.     t1 = time.time()
  162.     det_boxes,scores,ids = infer_image(net,img0,model_h,model_w,thred_nms=0.4,thred_cond=0.5)
  163.     t2 = time.time()
  164.     
  165.     print("cost time  %.2fs"%(t2-t1))
  166.     
  167.     for box,score,id in zip(det_boxes,scores,ids):
  168.         label = '%s:%.2f'%(dic_labels[id],score)
  169.         
  170.         plot_one_box(box.astype(np.int16), img0, color=(255,0,0), label=label, line_thickness=None)
  171.     cv2.imshow('img',img0)
  172.  
  173.     cv2.waitKey(0)

         下面是视频检测代码，我采用的是树莓派自带的摄像头，如果是其他深度学习摄像头，以下代码需要更改。

  1. # 视频检测
  2. import cv2
  3. import numpy as np
  4. import time
  5. from threading import Thread
  6.  
  7. def plot_one_box(x, img, color=None, label=None, line_thickness=None):
  8.     """
  9.     description: Plots one bounding box on image img,
  10.                  this function comes from YoLov5 project.
  11.     param: 
  12.         x:      a box likes [x1,y1,x2,y2]
  13.         img:    a opencv image object
  14.         color:  color to draw rectangle, such as (0,255,0)
  15.         label:  str
  16.         line_thickness: int
  17.     return:
  18.         no return
  19.     """
  20.     tl = (
  21.         line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1
  22.     )  # line/font thickness
  23.     color = color or [random.randint(0, 255) for _ in range(3)]
  24.     c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
  25.     cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
  26.     if label:
  27.         tf = max(tl - 1, 1)  # font thickness
  28.         t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
  29.         c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
  30.         cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA)  # filled
  31.         cv2.putText(
  32.             img,
  33.             label,
  34.             (c1[0], c1[1] - 2),
  35.             0,
  36.             tl / 3,
  37.             [225, 255, 255],
  38.             thickness=tf,
  39.             lineType=cv2.LINE_AA,
  40.         )
  41.  
  42. def post_process_opencv(outputs,model_h,model_w,img_h,img_w,thred_nms,thred_cond):
  43.     
  44.     conf = outputs[:,4].tolist()
  45.     c_x = outputs[:,0]/model_w*img_w
  46.     c_y = outputs[:,1]/model_h*img_h
  47.     w  = outputs[:,2]/model_w*img_w
  48.     h  = outputs[:,3]/model_h*img_h
  49.     p_cls = outputs[:,5:]
  50.     if len(p_cls.shape)==1:
  51.         p_cls = np.expand_dims(p_cls,1)
  52.     cls_id = np.argmax(p_cls,axis=1)
  53.  
  54.     p_x1 = np.expand_dims(c_x-w/2,-1)
  55.     p_y1 = np.expand_dims(c_y-h/2,-1)
  56.     p_x2 = np.expand_dims(c_x+w/2,-1)
  57.     p_y2 = np.expand_dims(c_y+h/2,-1)
  58.     areas = np.concatenate((p_x1,p_y1,p_x2,p_y2),axis=-1)
  59.     # print(areas.shape) 
  60.     areas = areas.tolist()
  61.     ids = cv2.dnn.NMSBoxes(areas,conf,thred_cond,thred_nms)
  62.     if len(ids)>0:
  63.         return  np.array(areas)[ids],np.array(conf)[ids],cls_id[ids]
  64.     else:
  65.         return [],[],[]
  66.  
  67. def infer_image(net,img0,model_h,model_w,thred_nms=0.4,thred_cond=0.5):
  68.  
  69.     img = img0.copy()
  70.     img = cv2.resize(img,[model_h,model_w])
  71.     blob = cv2.dnn.blobFromImage(img, scalefactor=1/255.0, swapRB=True)
  72.     net.setInput(blob)
  73.     outs = net.forward()[0]
  74.     
  75.     det_boxes,scores,ids = post_process_opencv(outs,model_h,model_w,img0.shape[0],img0.shape[1],thred_nms,thred_cond)
  76.     return det_boxes,scores,ids
  77.    
  78. global det_boxes_show
  79.  
  80. global scores_show
  81.  
  82. global ids_show
  83.  
  84. global FPS_show
  85.  
  86.  
  87. def m_detection(net,cap,model_h,model_w):
  88.     global det_boxes_show
  89.     global scores_show
  90.     global ids_show
  91.     global FPS_show
  92.     while True:
  93.         success, img0 = cap.read()
  94.         if success:
  95.  
  96.             t1 = time.time()
  97.             det_boxes,scores,ids = infer_image(net,img0,model_h,model_w,thred_nms=0.4,thred_cond=0.4)
  98.             t2 = time.time()
  99.             str_fps = "FPS: %.2f"%(1./(t2-t1))
  100.             
  101.             
  102.             det_boxes_show = det_boxes
  103.             scores_show = scores
  104.             ids_show = ids
  105.             FPS_show = str_fps
  106.             
  107.             # time.sleep(5)
  108.  
  109. if __name__=="__main__":
  110.     dic_labels = {0: 'person',
  111.                   1: 'bicycle',
  112.                   2: 'car',
  113.                   4: 'airplane',
  114.                   5: 'bus',
  115.                   6: 'train',
  116.                   7: 'truck',
  117.                   8: 'boat',
  118.                   9: 'traffic light',
  119.                   10: 'fire hydrant',
  120.                   11: 'stop sign',
  121.                   12: 'parking meter',
  122.                   13: 'bench',
  123.                   14: 'bird',
  124.                   15: 'cat',
  125.                   16: 'dog',
  126.                   17: 'horse',
  127.                   18: 'sheep',
  128.                   19: 'cow',
  129.                   20: 'elephant',
  130.                   21: 'bear',
  131.                   22: 'zebra',
  132.                   23: 'giraffe',
  133.                   24: 'backpack',
  134.                   25: 'umbrella',
  135.                   26: 'handbag',
  136.                   27: 'tie',
  137.                   28: 'suitcase',
  138.                   29: 'frisbee',
  139.                   30: 'skis',
  140.                   31: 'snowboard',
  141.                   32: 'sports ball',
  142.                   33: 'kite',
  143.                   34: 'baseball bat',
  144.                   35: 'baseball glove',
  145.                   36: 'skateboard',
  146.                   37: 'surfboard',
  147.                   38: 'tennis racket',
  148.                   39: 'bottle',
  149.                   40: 'wine glass',
  150.                   41: 'cup',
  151.                   42: 'fork',
  152.                   43: 'knife',
  153.                   44: 'spoon',
  154.                   45: 'bowl',
  155.                   46: 'banana',
  156.                   47: 'apple',
  157.                   48: 'sandwich',
  158.                   49: 'orange',
  159.                   50: 'broccoli',
  160.                   51: 'carrot',
  161.                   52: 'hot dog',
  162.                   53: 'pizza',
  163.                   54: 'donut',
  164.                   55: 'cake',
  165.                   56: 'chair',
  166.                   57: 'couch',
  167.                   58: 'potted plant',
  168.                   59: 'bed',
  169.                   60: 'dining table',
  170.                   61: 'toilet',
  171.                   62: 'tv',
  172.                   63: 'laptop',
  173.                   64: 'mouse',
  174.                   65: 'remote',
  175.                   66: 'keyboard',
  176.                   67: 'cell phone',
  177.                   68: 'microwave',
  178.                   69: 'oven',
  179.                   70: 'toaster',
  180.                   71: 'sink',
  181.                   72: 'refrigerator',
  182.                   73: 'book',
  183.                   74: 'clock',
  184.                   75: 'vase',
  185.                   76: 'scissors',
  186.                   77: 'teddy bear',
  187.                   78: 'hair drier',
  188.                   79: 'toothbrush'}
  189.  
  190.     model_h = 640
  191.     model_w = 640
  192.     file_model = 'yolov5n.onnx'
  193.     net = cv2.dnn.readNet(file_model)
  194.     
  195.     video = 0
  196.     cap = cv2.VideoCapture(video)
  197.     
  198.     m_thread = Thread(target=m_detection, args=([net,cap,model_h,model_w]),daemon=True)
  199.     m_thread.start()
  200.     
  201.     global det_boxes_show
  202.     global scores_show
  203.     global ids_show
  204.     global FPS_show
  205.     
  206.     
  207.     
  208.     det_boxes_show = []
  209.  
  210.     scores_show = []
  211.  
  212.     ids_show  =[]
  213.  
  214.     FPS_show = ""
  215.     
  216.     while True:
  217.         success, img0 = cap.read()
  218.         if success:
  219.             
  220.             for box,score,id in zip(det_boxes_show,scores_show,ids_show):
  221.                 label = '%s:%.2f'%(dic_labels[id],score)
  222.                 plot_one_box(box, img0, color=(255,0,0), label=label, line_thickness=None)
  223.                 
  224.             str_FPS = FPS_show
  225.             
  226.             cv2.putText(img0,str_FPS,(50,50),cv2.FONT_HERSHEY_COMPLEX,1,(0,255,0),3)
  227.             
  228.             
  229.             cv2.imshow("video",img0)
  230.         if cv2.waitKey(1) & 0xFF == ord('q'):
  231.             break
  232.     cap.release()

4. 结语

总体来说，树莓派实现YOLOv5的检测还是比较吃力的，后续可能会试试其他边缘设备，上述仅供参考，欢迎大家多多点赞交流。

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/AllinToyou/article/detail/147320
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号