erikbohnsack/pmbm

Python implementation of Poisson Multi-Bernoulli Mixture Filter for Multi-Object Tracking.

pmbm's Issues

Is it a real pmbm algorithms?

Hi, @erikbohnsack . thanks for supply this code. However, I confused this algorithm is pmbm ? because I found the code about possion birth is not valid. I think this code is mbm algorithm, is it rightly?

Undetected objects misdirected again

Hi, thanks for sharing this great code. Im unsure whether its a mistake or i dont understand something. Does this correspond to the update of the undetected objects modeled by Poisson point process that are missdetected again? Shouldn't it have a factor 1 - self.detection_probability then? My reference is the handout 5 from the edX course "Multi-Object Tracking for Automotive Systems" that describes the PMBM update.

pmbm/pmbm/pmbm.py

Lines 137 to 140 in 9fc4dbb

	# Update current tracks
	for target_id, target in self.targets.items():
	target.update(measurements, classes, self.current_time, self.filt)
	self.poisson.update(self.detection_probability)

how can i install murty?

i have tried the command pip3 install ./murty and it is not work. how can i solve the problem?
thx

How can I obtain ./Murty file?

When doing the command git submodule update, I cannot obtain the ./Murty file.

how to run this project?

Hello, thanks for your wonderful work and it's so kind of you to open source.

When I try to run your code by putting all codes in runforrest.ipynb file together, I found it may miss some files, such as the mot/fafe/showroom/showroom_all/showroom_bev_NN/logs/log-seq specified by log_root variable and the mot/fafe/showroom/weights_2019-05-03_14-01_epoch_110_fafe file specified by showroom_fafe variable.

Could you offer some help and point out some clues to run your code to get the results like yours.

Thanks a lot.

Question about the derivation of the CV/CA model

pmbm/utils/motion_models.py

Lines 96 to 99 in 9fc4dbb

	self.Q = np.array([[dt ** 3 / 3, 0, dt ** 2 / 2, 0],
	[0, dt ** 3 / 3, 0, dt ** 2 / 2],
	[dt ** 2 / 2, 0, dt, 0],
	[0, dt ** 2 / 2, 0, dt]]) * self.q

pmbm/utils/motion_models.py

Lines 214 to 219 in 9fc4dbb

	self.Q = np.array([[dt ** 5 / 20, 0, dt ** 4 / 8, 0, dt ** 3 / 6, 0],
	[0, dt ** 5 / 20, 0, dt ** 4 / 8, 0, dt ** 3 / 6],
	[dt ** 4 / 8, 0, dt ** 3 / 3, 0, dt ** 2 / 2, 0],
	[0, dt ** 4 / 8, 0, dt ** 3 / 3, 0, dt ** 2 / 2],
	[dt ** 3 / 6, 0, dt ** 2 / 2, 0, dt, 0],
	[0, dt ** 3 / 6, 0, dt ** 2 / 2, 0, dt]]) * self.q

Hi,
I'm confused about the form of the covariancce matrix Q in the CV/CA model, how to derivate it?

coord_transfer.py may have something wrong?

Hi,
When I read your code in detail, I'm confused about the calculation process in utils/coord_transfer.py:

def coordinate_transform_CV(state, imu_data, dt, object_class=None):
    assert state.shape == (4, 1)
    # Minus as the imu_data is for the ego vehicle
    angle = - imu_data.ru * dt

    # minus angle holds here as well
    rotation_matrix = np.array([[math.cos(angle), - math.sin(angle)],
                                [math.sin(angle), math.cos(angle)]])
    translation = np.array([[- imu_data.vl * dt],
                            [imu_data.vf * dt]])

    output_state = np.copy(state)
    output_state[0:2] = rotation_matrix @ state[0:2] - translation
    output_state[2:4] = rotation_matrix @ state[2:4]
    # same argument for minus translation here as above.
    return output_state

I understand that you use the IMU information to transform the objects state from last frame to current frame in ego vehicle, so we can get the absolute tracking velocity. So the equation

output_state[0:2] = rotation_matrix @ state[0:2] - translation

should be

output_state[0:2] = rotation_matrix @ (state[0:2] - translation)

when do the inverse transformation, right?

However when I correct it, the results seems the same. I'm not sure whether I'm wrong.