Hi Marjan Ramin

A classical way of tuning FP vs. FN is by using ROC curves. Say your classification is a function of a threshold (above the threshold you accept it, below it you reject it). By increasing the threshold your classifier becomes more specific and less sensitive (you reduce FP but increase FN). If you decrease the threshold your classifier becomes less specific and more sensitive (you reduce FN, but increase FP). The ROC curve will show you how your classifier behaves for all choices of the threshold values, then you choose in which point you want to operate. Of course normally there are no 'free lunches', i.e., by improving FP rate you 'pay the price' of a worse FN rate. Hope this is helpful.

Hello,

One thing you can do, that you have to compute  the PDF of tracking object and then measure..

Hello Marjan Ramin,

Your problem can have different solution. First, try to use other indicators of classification algorithm performance such as precision, recall, f1 score or cohen's kappa. The kappa coefficient can be high with a low accuracy. This means that your algorithm is good. Then, the training stage of your algorithm has to be correctly done. For example, if you use ANN, overfitting phenomenon can influence your results. So you have to use a number of training data less important than your current number of training data. 

Another important thing is your features. May be you have to change the extraction features method in order to get more discriminant attributes. Last thing, look the performances for each classes. May be you have 1 or 2 classes, which increase your FP and FN.

And try other classification algorithms. For example, in my study case, the best algorithm is k-nn. And I compared C4.5, CART, Multi-Class SVM, ANN, ANFIS. 

Thanks for all the answers. I am using DeepPed for detection part, and I guess it is accurate enough in feature extraction and detecting pedestrians.

I think the problem is that there are some differences between my results and ground truth, and these are not necessarily wrong detections. For example, there are some pedestrians in my results which are not annotated in my ground truth.

Ok, so, you have in a first time to delete the data (pedestrians) that are not in you ground truth. Then, check the results of your classification. Then, include the other pedestrians and re-check the detection outcomes. And finally, add new classes describing the actions (activities) of your pedestrians, which are not in your ground truth. For example, these classes can be: 'the unknown pedestrian turns to the right'. It is just an example. But all the new classes are defined by unknown pedestrian. 

These information help you? had I been clear enough?

Hello Marjan,

The best way to reduce false tracks in a tracking system is not to start them in the first place. This means very strict control of when to initiate a track. The problem was studied by Morefield in the late 1970s and he proposed an integer programming solution. A simpler solution like M out of N logic can work OK in practice but you must make some assumptions about the initial state (e.g. velocity restrictions). Have a look at my paper on HMM tracking, this has a very low false track rate on high false alarm density (sonar) data.

Morefield's paper is referenced in the overview paper below.

Conference Paper A High Performance 1-D Hidden Markov Model Tracker for Passi...

Article Taxonomy of multiple target tracking methods

I am trying to think out of the box way.May be project Green eye considered optical sensors to pixels or colour to extrated features in high speed as per CNN.

What about Voxcel or cross validation with Fusion sensors?