My initial goal with this project is a tracker with reidentification cabilities to track the target person. I've developed this tracker to put into a human tracking robot.

I've made a tracker only using orb features, a detector and a tracker. HMM...

• First the tracker is initialized with target features. If the target features are given, the system starts tracking else it starts collecting features from the center person. In my case the target should turn 360 degrees so it can be retracked from all angles. Only features from the upper portion of the detection are collected.
• After initialization the target features are found in the current frame. Then optical flow is initialized to track found features.
• Every nth frame the tracked points are populated by finding target features in the current frame to keep tracking continuous.
• If tracked points are below a certain threshold, retracking is actived.
• Every nth frame new features are added to the target features.
• Every nth frame target features and environment features are compared, if there is a match, matches are removed from the target features to reduce false positives.
• If there is no tracked points all detections are considered to find target person so removing false positives is necessary step to eliminate wrong tracking.

• OpenCV
• scipy
• sklearn
• numpy

• MobileNet SSD Repository
• You don't need to compile the MobileNet SSD, OpenCV DNN is used.
• I've used mobilenet_iter_73000.caffemodel from the original repository
• Put mobilenet_iter_73000.caffemodel into mobilenetssd folder.
• You can also use --mobilenet_model argument to pass the model path.

• Compile darknet with shared library.
• Put the library into the lib folder.
• You can also use DARKNET_LIB environment variable to pass the path.
• Weights, Config File and Data files should be downloaded.
• The detector only uses detections with the person tag

• imutils

python tracker.py

• Usage with MobileNet SSD Detector and 1000 Collected Features to start tracking

python tracker.py --detector mobilenet --mobilenet_model ./mobilenetssd/mobilenet_iter_73000.caffemodel --mobilenet_prototxt ./mobilenetssd/deploy.prototxt --n_tracked 1000 

• Usage with video input

python tracker.py --input video.avi

• Usage with YOLO Detector

DARKNET_LIB=./lib/darknet.so python tracker.py --detector yolo --yolo_cfg ./yolo/files/yolov3-tiny.cfg --yolo_weights ./yolo/files/yolov3-tiny.weights --yolo_meta ./yolo/files/coco.data

• To see other options you can use

python tracker.py --help

You can use ESC (if you didn't use with --no_gui) to start tracking earlier without reaching n_tracked number of features.

from options import args
from tracker import Tracker

class MyTracker(Tracker):
	def __init__(self, args):
		# Initiliaze Tracker class
		Tracker.__init__(self, args)

		self._cap = cv2.VideoCapture(self.args.input)
		# Camera Settings
		self._cap.set(cv2.CAP_PROP_FRAME_WIDTH, self.args.camera_width)
		self._cap.set(cv2.CAP_PROP_FRAME_HEIGHT, self.args.camera_height)
		if (self.args.camera_fps):
			self._cap.set(cv2.CAP_PROP_FPS, self.args.camera_fps)

	def get_frame(self):
		ret, self.frame = self._cap.read()
		return ret, self.frame
		
	# You can use this function to interract with the tracked points.
	# For example publish points with a ROS Node
	def output_function(self):
		print('Tracked Points: %d' % len(self.track_points))
		# Return False if you want to close the system
		return True
		
	# This function executed at the end of each collect_features function
	# Useful when there is no OpenCV window
	def collection_output(self):
		print('Collected Features: %d' % len(self.tkps))
		
	# This function executed while closing.
	# Useful to set certain events in case of threading.
	# Check zeromq_example.py for an example
	def finish(self):
		pass

tracker = MyTracker(args)
tracker.initiliaze_target()
tracker.run()

NOTE: By default it is using cv2.VideoCapture(self.args.input), so no need to initiliaze it. This code is just for the example.

• example.py Minimal example

• ros_example.py shows how to integrate the library into ROS.

• zeromq_example.py shows how to use it with ZMQ Library.

• All examples can be executed, if you want to test the ROS example you can use ros_camera.py to publish images from the camera.