I suggest you techniques of blockmathcing. They are fast and robust. Here, there is an of may papaer that tries to solve this problem

Article Block matching algorithm for motion estimation based on Arti...

In similar cases, I would apply some prepossessing filters and techniques (e.g. Gaussian Mixture Models) to the raw image/frame before feeding the image to the object detection algorithm.

Also, see this paper: "Dehazing for Image and Video Using Guided Filter."

http://www.scirp.org/journal/PaperDownload.aspx?DOI=10.4236/ojapps.2012.24B030

Dear Aydin Ayanzadeh ,

provided that you have a working object detection algorithm to detect objects in individual frames (which can be a lot of things, but I would suggest you employ a CNN such as YOLO), there are several standard Visual Tracking approaches that have been developed in the last 10 years that are used to track these objects.

There is, however, a method called Deep Sort that has becoming widely used. It is officially called Simple Online and Realtime Tracking (SORT) and is a pragmatic approach to multiple object tracking with a focus on simple, effective algorithms. Here are the references:

• https://github.com/nwojke/deep_sort The implementation here extends the original SORT algorithm to integrate appearance information based on a deep appearance descriptor
• Simple Online and Realtime Tracking with a Deep Association Metric - https://arxiv.org/abs/1703.07402

Please tell me if it helped.

Aldo

Aldo Von Wangenheim , Abdulrahman Alruban Erik Cuevas Thanks for your response i want to share couple of frames with you to get more accurate response. There are two type of datasets:

• very low resultion images: (Fish in silty environments):

Aldo Von Wangenheim Abdulrahman Alruban Erik Cuevas Moreover, We have another dataset which has the better quality and contrast.

• The region of Interest has low Contrast.

Dear Aydin Ayanzadeh ,

you have scenes that are indeed very difficult ones. I cannot imagine that you could detect the fish in such images using anything other than Object Detection CNNs.

Before you endeavor implementing fish tracking, I would try to reliably identify the fish. The rest will be trivial, since it seems you won't have swarms where you have to identify individual fish.

For newbies in CNN, there are several tutorials on Object Detection. You have basically three APIs you could use:

• TensorFlow (Google's flagship) - low-level, full of ties, several complicated installation requirements and complicated to use;
• Keras (easy to use but with lots of bugs);
• PyTorch/fast.ai - the new trend in the scientific community.

I'll provide you tutorials for both TensorFlow and PyTorch:

TF (using the high-level TF Object Detection API):

• https://www.edureka.co/blog/tensorflow-object-detection-tutorial/
• https://www.datacamp.com/community/tutorials/object-detection-guide
• https://pythonprogramming.net/introduction-use-tensorflow-object-detection-api-tutorial/

PyTorch (my suggestion!):

• https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection
• https://towardsdatascience.com/object-detection-and-tracking-in-pytorch-b3cf1a696a98
• https://medium.com/@smallfishbigsea/understand-ssd-and-implement-your-own-caa3232cd6ad

If you want to use the very high level fast.ai API (based upon PyTorch), here's a video-tutorial:

• https://www.youtube.com/watch?v=Z0ssNAbe81M

Please let me know if this helped.

Aldo

Check if this can work

Walsh-Hadamard Kernel based Particle Filter work is recently reported in https://ieeexplore.ieee.org/document/8815857