not tracking #1
Comments
|
Could you provide more detail (error messages or images?)? I assume that you ran the main.py. I ran the example again. It seems to work fine. There could be a few occations that the detector could miss detecting the car. In these cases, the most recently updated bounding box would be used for tracking (e.g., Frame 5). Was this the situation that you encountered? Here are a few examples from my end:
Let me know if this would help. Best, |
|
Ya that is the case. But I run the tracker.py not the main.py.
…On Sun, Feb 25, 2018 at 7:59 AM, Kyle Guan ***@***.***> wrote:
Could you provide more detail (error messages or images?)? I assume that
you ran the main.py. I ran the example again. It seems to work fine. There
could be a few occations that the detector could miss detecting the car. In
these cases, the most recently updated bounding box would be used (e.g.,
Frame 5). Was this the situation that you encountered?
Here are a few examples from my end:
1. measurement and update for frame 4, respectively.
[image: frame4_measurement]
<https://user-images.githubusercontent.com/12797392/36643483-05b9d188-1a1a-11e8-958a-d3e8fc423525.png>
[image: frame4_update]
<https://user-images.githubusercontent.com/12797392/36643484-05c42304-1a1a-11e8-8821-4df63bda7775.png>
2. measurement and update for Frame 8
[image: frame8_measurement]
<https://user-images.githubusercontent.com/12797392/36643512-4cb5bed0-1a1a-11e8-8720-d39adb271fb8.png>
[image: frame8_update]
<https://user-images.githubusercontent.com/12797392/36643515-55c30cee-1a1a-11e8-96b2-06e9ebc60c8d.png>
Let me know if this would help.
Best,
Kyle
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdSqXk4aGCp06xz697Gtg_YYG6A3Uks5tYYNhgaJpZM4SSS9P>
.
|
|
By the way The tensorflow detection api with the SSD.Mobilet architecture
is not accurate than the pure SSD detector in keras. I tested it in some
images and the score is very less for this API.....Why we use it? do you
have SSD with a tracker like kalman filter? or do you know any?
thank you
On Sun, Feb 25, 2018 at 9:27 PM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
… Ya that is the case. But I run the tracker.py not the main.py.
On Sun, Feb 25, 2018 at 7:59 AM, Kyle Guan ***@***.***>
wrote:
> Could you provide more detail (error messages or images?)? I assume that
> you ran the main.py. I ran the example again. It seems to work fine. There
> could be a few occations that the detector could miss detecting the car. In
> these cases, the most recently updated bounding box would be used (e.g.,
> Frame 5). Was this the situation that you encountered?
> Here are a few examples from my end:
>
> 1. measurement and update for frame 4, respectively.
> [image: frame4_measurement]
> <https://user-images.githubusercontent.com/12797392/36643483-05b9d188-1a1a-11e8-958a-d3e8fc423525.png>
> [image: frame4_update]
> <https://user-images.githubusercontent.com/12797392/36643484-05c42304-1a1a-11e8-8821-4df63bda7775.png>
> 2. measurement and update for Frame 8
> [image: frame8_measurement]
> <https://user-images.githubusercontent.com/12797392/36643512-4cb5bed0-1a1a-11e8-8720-d39adb271fb8.png>
> [image: frame8_update]
> <https://user-images.githubusercontent.com/12797392/36643515-55c30cee-1a1a-11e8-96b2-06e9ebc60c8d.png>
>
> Let me know if this would help.
>
> Best,
> Kyle
>
> —
> You are receiving this because you authored the thread.
> Reply to this email directly, view it on GitHub
> <#1 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AgXgdSqXk4aGCp06xz697Gtg_YYG6A3Uks5tYYNhgaJpZM4SSS9P>
> .
>
|
|
[image: Inline image 1]
On Sun, Feb 25, 2018 at 9:29 PM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
… By the way The tensorflow detection api with the SSD.Mobilet architecture
is not accurate than the pure SSD detector in keras. I tested it in some
images and the score is very less for this API.....Why we use it? do you
have SSD with a tracker like kalman filter? or do you know any?
thank you
On Sun, Feb 25, 2018 at 9:27 PM, Amare Mahtsentu ***@***.***>
wrote:
> Ya that is the case. But I run the tracker.py not the main.py.
>
> On Sun, Feb 25, 2018 at 7:59 AM, Kyle Guan ***@***.***>
> wrote:
>
>> Could you provide more detail (error messages or images?)? I assume that
>> you ran the main.py. I ran the example again. It seems to work fine. There
>> could be a few occations that the detector could miss detecting the car. In
>> these cases, the most recently updated bounding box would be used (e.g.,
>> Frame 5). Was this the situation that you encountered?
>> Here are a few examples from my end:
>>
>> 1. measurement and update for frame 4, respectively.
>> [image: frame4_measurement]
>> <https://user-images.githubusercontent.com/12797392/36643483-05b9d188-1a1a-11e8-958a-d3e8fc423525.png>
>> [image: frame4_update]
>> <https://user-images.githubusercontent.com/12797392/36643484-05c42304-1a1a-11e8-8821-4df63bda7775.png>
>> 2. measurement and update for Frame 8
>> [image: frame8_measurement]
>> <https://user-images.githubusercontent.com/12797392/36643512-4cb5bed0-1a1a-11e8-8720-d39adb271fb8.png>
>> [image: frame8_update]
>> <https://user-images.githubusercontent.com/12797392/36643515-55c30cee-1a1a-11e8-96b2-06e9ebc60c8d.png>
>>
>> Let me know if this would help.
>>
>> Best,
>> Kyle
>>
>> —
>> You are receiving this because you authored the thread.
>> Reply to this email directly, view it on GitHub
>> <#1 (comment)>,
>> or mute the thread
>> <https://github.com/notifications/unsubscribe-auth/AgXgdSqXk4aGCp06xz697Gtg_YYG6A3Uks5tYYNhgaJpZM4SSS9P>
>> .
>>
>
>
|
|
|
|
@3073 , thanks for your comments and feedback.
Hope these could be helpful. Best, Kyle |
|
Thank you for real. you are a smart developer with satisfying answer for
you customers.
what I am saying is do you know any github code that illustrates. SSD
detection and any kind of tracking algorithm for the detected objects like
in the following link.
https://github.com/ykamikawa/yolov2
In this link we are using YOLO for detection and kalman filter for tracking
the detected objects. But since YOLO is difficult to understand I need to
adapt SSD. That is my question.
Sorry if I kill your time. Thank you very very much.
one more thing: if you have time and able to do that for me I will
definitely donate for your time.
…On Mon, Feb 26, 2018 at 8:53 AM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073> , thanks for your comments and feedback.
1.
The code in tracker.py is meant to quickly illustrate the impact noise
covariance matrix of an measurement. You can modify the code:
127 images = [plt.imread(file) for file in glob.glob('./test_images/*.
jpg')]
128 img=images[3]
to suit your purpose.
2.
I agree with you that the detection (localization) quality of
SSD-Mobilenet architecture is far from perfect. I use it mainly because the
fast inference time (<200 ms on a CPU only laptop). Of course, you can use
other detection architectures/APIs, subject to the trade-off between
accuracy and inference time. In my experience, there will always be cases
that a general-purpose object detection API could fail. Thus I plan to
optimize SSD-Mobilenet for vehicle detection only with new training images.
I will keep you posted on the progress.
3.
I am not quite sure about your question "do you have SSD with a
tracker like kalman filter". From my experience, SSD only provides the
measurement part (by outputing the bounding box in a frame) for Bayesian
filtering (could be KF, UKF, or EKF). I don't believe SSD, in its most
basic design, has a built-in predict and update processes. I could be wrong.
Hope these could be helpful.
Best, Kyle
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdWHFCXFreaZ-66UfUEtJjBLQdxzSks5tYuGMgaJpZM4SSS9P>
.
|
|
@3073 , I am only aware of Simple, online, and realtime tracking (SORT) for multi-object tracking in this repo: https://github.com/abewley/sort. Also if you want to dig deep into the concepts/techniques of varios Bayesian filters for tracking, the following repo is a very good source: https://github.com/rlabbe/Kalman-and-Bayesian-Filters-in-Python Hope these would help. Best, |
|
the link below is the perfect example that I need . but I do not know
what type of detector is used in detector.py
https://github.com/ZidanMusk/experimenting-with-sort
anyways thank you for your time
…On Mon, Feb 26, 2018 at 3:55 PM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073> , I am only aware of Simple, online, and
realtime tracking (SORT) for multi-object tracking in this repo:
https://github.com/abewley/sort. Also if you want to dig deep into the
concepts/techniques of varios Bayesian filters for tracking, the following
repo is a very good source: https://github.com/rlabbe/
Kalman-and-Bayesian-Filters-in-Python
Hope these would help.
Best,
Kyle
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdUJYCvZYhAJm-ANFpvcKiPZj5NYZks5tY0RygaJpZM4SSS9P>
.
|
|
I have no idea either. The best way is to contact the author/developer of the repo directly. |
|
Thank you
I have tested your gift (
https://github.com/kcg2015/Vehicle-Detection-and-Tracking) and It is
working for a video file with minor modification to suit my need.
could you help me how can I use it Webcam?
…On Tue, Feb 27, 2018 at 2:33 PM, Kyle Guan ***@***.***> wrote:
I have no idea either. The best way is to contact the author/developer of
the repo directly.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdY7-AYdnZMeLMwHmXsU2Q-2OD7cdks5tZIK7gaJpZM4SSS9P>
.
|
|
I have not worked on webcam capturing yet. You can check out this YouTube video and the code for more information. https://www.youtube.com/watch?v=MyAOtvwTkT0. Hope this would help. |
|
Thank you.
do you have some time to show me How the Main.py works in your repo?
I did not understand the codes in:
# Deal with matched detection
# Deal with unmatched detection
# Deal with unmatched tracks
…On Wed, Feb 28, 2018 at 8:24 PM, Kyle Guan ***@***.***> wrote:
I have not worked on webcam capturing yet. You can check out this YouTube
video and the code for more information. https://www.youtube.com/watch?
v=MyAOtvwTkT0. Hope this would help.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdeeJ4pp53BnKMOKmC9QJ-NfTLCCKks5tZ3gBgaJpZM4SSS9P>
.
|
|
@3073 , please give me some time to update the Readme file of the repo. I will provide more details about these three code blocks. |
|
Thank you
take your time my friend
…On Thu, Mar 1, 2018 at 4:43 PM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073> , please give me some time to update the
Readme file of the repo. I will provide more details about these three code
blocks.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdUnnJCCnrPD1zCia3X99C1ZHMGXaks5taJXAgaJpZM4SSS9P>
.
|
|
Is there a possibility to use this for counting the tracked persons. How
that can be possible
thank you
On Fri, Mar 2, 2018 at 2:40 AM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
… Thank you
take your time my friend
On Thu, Mar 1, 2018 at 4:43 PM, Kyle Guan ***@***.***>
wrote:
> @3073 <https://github.com/3073> , please give me some time to update the
> Readme file of the repo. I will provide more details about these three code
> blocks.
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#1 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AgXgdUnnJCCnrPD1zCia3X99C1ZHMGXaks5taJXAgaJpZM4SSS9P>
> .
>
|
|
Yes. The number of tracked persons equals to matched detections. |
|
Thank you
I am really sorry that I make you busy by these too many question.
Normally I am studying this to count the number of people entering a bus.
do you have any idea that a tracked object will not be counted twice?
…On Sun, Mar 4, 2018 at 6:28 PM, Kyle Guan ***@***.***> wrote:
Yes. The number of tracked persons equals to matched detections.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdVIInjggOYIn7D9MnYyQo8Oh8_-Qks5tbKLfgaJpZM4SSS9P>
.
|
|
that means how to avoid counting of a person in the remaining frames if it
is detected and tracked in the first frame?
On Mon, Mar 5, 2018 at 2:15 AM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
…
Thank you
I am really sorry that I make you busy by these too many question.
Normally I am studying this to count the number of people entering a bus.
do you have any idea that a tracked object will not be counted twice?
On Sun, Mar 4, 2018 at 6:28 PM, Kyle Guan ***@***.***>
wrote:
> Yes. The number of tracked persons equals to matched detections.
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#1 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AgXgdVIInjggOYIn7D9MnYyQo8Oh8_-Qks5tbKLfgaJpZM4SSS9P>
> .
>
|
|
@3073, no worry. Because the matching between the detection and the track is one-to-one, i.e, a detection can be matched to only (at most) one track, and a track can have only (at most) one associated detection, I doubt that a tracked detection will be double-counted. Does it sound reasonable to you? Let me know. |
|
by the way excuse me that I do not STAR your repo until now but you
answered me (in fact you teach me) many questions. I will do that before I
start my work!!!
Ya it sounds good and I will try my best and let you now the result.
thank you for real
…On Mon, Mar 5, 2018 at 6:13 AM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073>, no worry. Because the matching between
the detection and the track is one-to-one, i.e, a detection can be matched
to only (at most) one track, and a track can have only (at most) one
associated detection, I doubt any a tracked detection will be double
counted. Does it sound reasonable to you? Let me know.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdd_clSyVsgPROj4oYWE4NzMImk_1ks5tbUgDgaJpZM4SSS9P>
.
|
|
@3073 , I am glad this repo can be helpful to you. Thanks for the star. Again, please bear with me as it takes some time to update README file and some code blocks. This is not my day-job, so progress could be slow. |
|
ok. I stick with you
…On Mon, Mar 5, 2018 at 12:19 PM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073> , I am glad this repo can be helpful to
you. Thanks for the star. Again, please bear with me as it takes some time
to update README file and some code blocks. This is not my day-job, so
progress could be slow.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdT7_WouSknFoBMVoxN53GDYSEu4vks5tbZ3NgaJpZM4SSS9P>
.
|
|
Hi
I was testing it to count a person in a video.
I feed a video that contain a single same person in all the frames.
but it is counting the person many times ( equal to the number of frames
that contain the person) provide me some ideas
how do I identify the person in the second and third frame is the person
that is detected in the first frame?
Thank you very much
On Mon, Mar 5, 2018 at 12:21 PM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
… ok. I stick with you
On Mon, Mar 5, 2018 at 12:19 PM, Kyle Guan ***@***.***>
wrote:
> @3073 <https://github.com/3073> , I am glad this repo can be helpful to
> you. Thanks for the star. Again, please bear with me as it takes some time
> to update README file and some code blocks. This is not my day-job, so
> progress could be slow.
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#1 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AgXgdT7_WouSknFoBMVoxN53GDYSEu4vks5tbZ3NgaJpZM4SSS9P>
> .
>
|
|
Hi, @3073, I have never implementd object-counting feature before, so I have don't have any first-hand experience. So my guess is (still) counting the number of matched tracks (or detections) for a given frame. Also could you post your code (block) on github (public or private at your own choice)? Without going through your code, it is very hard to pinpoint the what happened. |
|
Ok thank you
…On Sun, Mar 11, 2018 at 1:24 PM, Kyle Guan ***@***.***> wrote:
Hi, @3073 <https://github.com/3073>, I have never implementd
object-counting feature before, so I have don't have any first-hand
experience. So my guess is (still) counting the number of matched tracks
(or detections) for a given frame. Also could you post your code (block) on
github (public or private at your own choice)? Without going through your
code, it is very hard to pinpoint the what happened.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdf4vZG9CN3DpVi5ADOqSRzkwUz2Iks5tdYf0gaJpZM4SSS9P>
.
|
|
Hi how are you. look below for my test on a video
what is wrong with the detection box that they double for single person?
On Sun, Mar 11, 2018 at 1:32 PM, Amare Mahtsentu <a.mahtsentu@gmail.com>
wrote:
… Ok thank you
On Sun, Mar 11, 2018 at 1:24 PM, Kyle Guan ***@***.***>
wrote:
> Hi, @3073 <https://github.com/3073>, I have never implementd
> object-counting feature before, so I have don't have any first-hand
> experience. So my guess is (still) counting the number of matched tracks
> (or detections) for a given frame. Also could you post your code (block) on
> github (public or private at your own choice)? Without going through your
> code, it is very hard to pinpoint the what happened.
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#1 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AgXgdf4vZG9CN3DpVi5ADOqSRzkwUz2Iks5tdYf0gaJpZM4SSS9P>
> .
>
|
|
@3073, could you show me an image of multiple boxes on the same object? |
|
here is the video
…On Tue, Mar 13, 2018 at 7:54 AM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073>, could you show me an image of multiple
boxes on the same object?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgde28aSQcpyhgY9o5yDsU1KgifzYMks5td921gaJpZM4SSS9P>
.
|
|
For some reason, I can not see your video. See if you can upload it on YouTube and send me a link? |
|
Dear my friend Deal with matched detectionDeal with unmatched detectionDeal with unmatched tracksLat time you promised to help me understand me all these thing. Thank you |
|
@3073 , I just update the readme file in the repo. It should be very clear the functionalities of these three code blocks. |
|
very nice |
|
Cool! I am going to close this issue. |
|
Hi there how are you....
I have this issue with the repo
I added dlib tracker with SSD mobile net and it is not detecting the
person. the detection box is not going in the direction the person is
moving rather the box is moving down ward.
ca you fix this....? i did not change the code in the detector.py
…On Thu, Apr 26, 2018 at 7:01 AM, Kyle Guan ***@***.***> wrote:
Cool! I am going to close this issue.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdeUPwq2EZoGzzUXTSz60CqPD1czBks5tsdMlgaJpZM4SSS9P>
.
|
|
@3073 , I am afraid that I can not be of much help to you. I am not familiar with dlib tracker at all. Also, from your description, it is very hard to pinpoint what the problem is let alone to fix it. All I can say is that if the performance of SSD Moblenet is not up to your task, consider using a different one in Tensorflow Object Dection API. |
|
Thank you very much
I have fixed it......
…On Thu, May 31, 2018 at 7:14 PM, Kyle Guan ***@***.***> wrote:
@3073 <https://github.com/3073> , I am afraid that I can not be of much
help to you. I am not familiar with dlib tracker at all. Also, from your
description, it is very hard to pinpoint what the problem is let alone how
to fix it. All I can say is that if the performance of SSD Moblenet is not
up to your task, consider using a different one in Tensorflow Object
Dection API.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AgXgdYcR6aPE_t0nZFei3QgSFUh0khfaks5t4KNzgaJpZM4SSS9P>
.
|
|
Hi how are you!!! have you check any possibility to run this code on raspberry pi? image classification on raspberry pi using SSD Moblenet ..... |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
the tacker is not tracking the sequence of images in "test_mages" folder.
the position of the bounding boxes do not change. it well track the "frame003" image
The text was updated successfully, but these errors were encountered: