Yolo v3 Object Detection Dash App

利用Dash poltly搭建的Yolo V3目标检测应用，web可视化检测结果。

Getting started

这里没有用Darknet，借助于Yolo-v3，利用Tensroflow构建的Yolo模型。不过TF的版本真的是最令人头痛的问题。
所以这里声明：

  1. tensorflow-gpu==2.0.0
  2. CUDA 10.0
  3. CUDNN 7.4.2

下载预训练权重

下载coco数据集预训练的权重。

  1. wget -P weights https://pjreddie.com/media/files/yolov3.weights

将权重转换为TF格式

Load the weights using load_weights.py script. This will convert the yolov3 weights into TensorFlow .ckpt model files!

  1. python load_weights.py

Running the model

You can run the model using detect.py script. The script works on images, video or your webcam. Don’t forget to set the IoU (Intersection over Union) and confidence thresholds.

Usage

  1. python detect.py <images/video/webcam> <iou threshold> <confidence threshold> <filenames>

Images example

Let’s run an example using sample images.

  1. python detect.py images 0.5 0.5 data/images/dog.jpg data/images/office.jpg

Then you can find the detections in the detections folder.

Video example

视频更改了部分指令：

  1. python detect.py video 0.5 0.5 data/video/shinjuku.mp4 outputfilename

这里遇到一个问题，opencv不能够编码X264. 而浏览器只能解析几个格式，这个问题还没有解决。

  1. # fourcc = cv2.VideoWriter_fourcc(*'X264')
  2. fourcc = cv2.VideoWriter_fourcc(*'mp4v')
  4. # fourcc = cv2.VideoWriter_fourcc(*'vp09')
  6. fps = cap.get(cv2.CAP_PROP_FPS)
  7. # out = cv2.VideoWriter('./assets/detections/{}.webm'.format(outputname[0]), fourcc, fps,
  8. #                       (int(frame_size[0]), int(frame_size[1])))
  9. out = cv2.VideoWriter('./assets/detections/{}.mp4'.format(outputname[0]), fourcc, fps,
  10. (int(frame_size[0]), int(frame_size[1])))
  1. 保存实时的视频流检测结果。
    在utils.py中添加自定义函数，将每个frame的检测结果和个数保存成字典，在detect .py while循环中将每一帧的结果保存下来。
  1. def statis_frame(frame, boxes_dicts, class_names,):
  2.     res = {}
  3.     boxes_dict = boxes_dicts[0]         # 检测结果 
  4.     for cls in range(len(class_names)): #80
  5.         boxes = boxes_dict[cls]
  6.         if np.size(boxes) != 0:
  7.             res[class_names[cls]]=len(boxes)
  9.     return res
  1. import pandas as pd 
  2. try:
  3.   res = []
  4.   while True:
  5.     ret, frame = cap.read()
  6.     if not ret:
  7.       break
  8.       resized_frame = cv2.resize(frame, dsize=_MODEL_SIZE[::-1],
  9.                                  interpolation=cv2.INTER_NEAREST)
  10.       detection_result = sess.run(detections,
  11.                                   feed_dict={inputs: [resized_frame]})
  13.       draw_frame(frame, frame_size, detection_result,
  14.                  class_names, _MODEL_SIZE)
  16.       res.append(statis_frame(frame,detection_result,class_names))
  17.       #                     cv2.imshow(win_name, frame)
  19.       key = cv2.waitKey(1) & 0xFF
  21.       if key == ord('q'):
  22.         break
  24.       out.write(frame)
  25.  finally:
  26.       pd.DataFrame(res).to_csv('res_csv/{}.csv'.format(outputname[0]))
  27.       cv2.destroyAllWindows()
  28.       cap.release()
  29.       print('Detections have been saved successfully.')

Webcam example

The script can also be ran using your laptops webcam as the input. Example command shown below.

  1. python detect.py webcam 0.5 0.5

The detections will be saved as ‘detections.mp4’ in the data/detections folder.

Dash App

  1. ![xF3Qid](https://cdn.jsdelivr.net/gh/Flionay/pic_bed@master/Upic/202104/xF3Qid.png)