alphapose

Tracking pedestrians with AlphaPose

An example of how to integrate Norfair into the video inference loop of a pre existing repository. This example uses Norfair to try out custom trackers on AlphaPose.

Instructions

1. Install Norfair with pip install norfair[video].

2. Follow the instructions to install the Pytorch version of AlphaPose.

3. Apply this diff to this commit on AlphaPose and use their video_demo.py to process your video.

   diff --git a/dataloader.py b/dataloader.py
   index ed6ee90..a7dedb0 100644
   --- a/dataloader.py
   +++ b/dataloader.py
   @@ -17,6 +17,8 @@ import cv2
   import json
   import numpy as np
   +import norfair
   import time
   import torch.multiprocessing as mp
   from multiprocessing import Process
   @@ -606,6 +608,17 @@ class WebcamLoader:
           # indicate that the thread should be stopped
           self.stopped = True
   
   +detection_threshold = 0.2
   +keypoint_dist_threshold = None
   +def keypoints_distance(detected_pose, tracked_pose):
   +    distances = np.linalg.norm(detected_pose.points - tracked_pose.estimate, axis=1)
   +    match_num = np.count_nonzero(
   +        (distances < keypoint_dist_threshold)
   +        * (detected_pose.scores > detection_threshold)
   +        * (tracked_pose.last_detection.scores > detection_threshold)
   +    )
   +    return 1 / (1 + match_num)
   +
   class DataWriter:
       def __init__(self, save_video=False,
                   savepath='demos/res/1.avi', fourcc=cv2.VideoWriter_fourcc(*'XVID'), fps=25, frameSize=(640,480),
   @@ -624,6 +637,11 @@ class DataWriter:
           if opt.save_img:
               if not os.path.exists(opt.outputpath + '/vis'):
                   os.mkdir(opt.outputpath + '/vis')
   +        self.tracker = norfair.Tracker(
   +            distance_function=keypoints_distance,
   +            distance_threshold=0.3,
   +            detection_threshold=0.2
   +        )
   
       def start(self):
           # start a thread to read frames from the file video stream
   @@ -672,7 +690,15 @@ class DataWriter:
                       }
                       self.final_result.append(result)
                       if opt.save_img or opt.save_video or opt.vis:
   -                        img = vis_frame(orig_img, result)
   +                        img = orig_img.copy()
   +                        global keypoint_dist_threshold
   +                        keypoint_dist_threshold = img.shape[0] / 30
   +                        detections = [
   +                            norfair.Detection(p['keypoints'].numpy(), scores=p['kp_score'].squeeze().numpy())
   +                            for p in result['result']
   +                        ]
   +                        tracked_objects = self.tracker.update(detections=detections)
   +                        norfair.draw_tracked_objects(img, tracked_objects)
                           if opt.vis:
                               cv2.imshow("AlphaPose Demo", img)
                               cv2.waitKey(30)

Explanation

With Norfair, you can try out your own custom tracker on the very accurate poses produced by AlphaPose by just integrating it into AlphaPose itself, and therefore avoiding the difficult job of decoupling the model from the code base.

This produces the following results: