What Model would best be suitable for my dataset?

Question

I am kind of overwhelmed with the amount of models there are so finding the one that best suits my dataset is proving kind of difficult. The Dataset I have is as follows

, its produced by a Radar, which outputs a row of values for a signal that it detects for a target.

Plotting across the row gives me the following wave, and as we go below the rows, we get the translation in the x-axis suggesting movement of the target, my dataset that I want to feed the model will have the following features: The average value of the two peaks for signal strength, the x-axis average value (multiplied by .77 meters), this will be the case for all of the waves for each target as each # of target will have these waves associated with it, tracking the movement of these waves shows the distance change, the change in amplitude and so on

Shows the translation of x-axis showing movement

I am currently working on a script that will try and get all the times when the radar detects something and get the amplitude column values and fill in the dataset, if I can't get the script to work I will just do it manually:

Link to what the Radar outputs

https://drive.google.com/file/d/1IJOebiXuScjLPytemulcXph7ZB1X65wU/view

The model that I will use will be different than what the radar outputs This is how I want to set the dataset for the model

Reference images to what I referred

https://i.sstatic.net/hZ1gp.jpg

I might also add another column that gets the average velocity of those points since we have time, distance given by the Radar. I will have a similar dataset for pedestrians as well, the first dataset was for vehicles, but there will be one for pedestrians as well, I want a model that can predict, once the training is done if the target was either a pedestrian or a vehicle given the features. What model would best fit this sort of data, a top 5 list would be super appreciated!

Thank You

Answer 1

From the data I suppose this is a timeseries problem and assuming you have continuous data, "since its a radar"

Generally the most used "best" approach is using stacked Long-Short Term Memory "LSTM" or some of its variants like GRU etc.. if you want to predict some specific cell in the future "e.g. row-column".. This might help get you started: https://keras.io/examples/timeseries/

If you are trying to predict the whole row, I would suggest using LSTM/variants AutoEncoders that would "supposedly" give you a better result in that case Thats also an introduction to that approach: https://towardsdatascience.com/step-by-step-understanding-lstm-autoencoder-layers-ffab055b6352

If your data is not continuous, I would suggest to use Convolutional Neural Networks "CNN" -AutoEncoders- to extract the features then feeding that to some fully connected layers to either outputs some categorization "e.g. distance from 1-5 or 5-10" or even the actual distance, but an important remark, this approach needs a lot of data points. it can be something like this: https://keras.io/examples/audio/

There are some research papers suggesting to convert your data points into 2D using fft or continuous wavelets then feeding that to a CNN to extract features from, but I haven't got any good results from that in different sets of problems.

P.S. Scaling the data beforehand "e.g. between the range of [0,1]" is essential in this case because apparently you have different units for each column.