I don't think missed detections and delayed detections are the same thing or handled in the same way. With delayed detections (e.g. due to data transfers) you can refine your current estimate by incorporating the delayed measurement when it eventually arrives.  

Please refer to:

Coordinated path-following in the presence of communication losses and time delays, Reza Ghabcheloo, António Aguiar, António Pascoal, Carlos Silvestre, I. Kaminer, J. Hespanha, SIAM - Journal on Control and Optimization, Vol. 48, No. 1, pp. 234-265, 2009 (http://welcome.isr.ist.utl.pt/img/pdfs/1977_SIAM08.pdf)

Path-following algorithms have been used for target tracking. See for example:

Moving path following for autonomous robotic vehicles

T. Oliveira, P. Encarnação, and A. Aguiar

Proceedings of ECC13 - 12th biannual European Control Conference, Zurich, Switzerland, July 2013, p. 3320-3325. (http://www.ucp.pt/site/resources/documents/Docente%20-%20PedroEncarnacao/ECC2013_MovingPathFollowing.pdf.zip(

target detection or path finding i think Direction of arrival DOA technology helps,

in that beam forming or beam estimation algorithms like MUSIC Algorthms , Barlette algorithms may suits .

Thanks for your answers, 

but as you know the Probability of Detection is one of the most important parameters in the design of PDAF/JPDAF filters. This phenomena causes that all of the target originated measurements sequences do not arrive to the tracker. Also Probability of detection arises due to the sensor resolution and signal processing on measurements. 

On the other hand, there are other parameters that affect the tracking quality, like delays and data losses. Sensors' faults and network quality cause some measurements delay and some of them lost.  

My question is that Can we treat data loss (due to network) and probability of detection (due to sensor features) as the same phenomena? 

Are there any approach or method that can handle tracking  or tolerate  in the presence of large data loss? 

I think the need for the Probability of Detection (PD) parameter is a big weakness of PDA and JPDA. PD depends on so many factors; it is highly variable and essentially unknowable, yet it has a big effect on the tracker behavior (e.g. the filter gain, and on track management, when the tracker is integrated with a track confidence model).

So I think you could justify using it to account for network loss/delays if it has the desired effect. Although, without knowing more about the architecture of your system, I'm not sure how you would go about doing this. For instance, ignoring network delays/loss, if the sensor sends single measurements with a (continuous) time stamp to the tracker when they are detected, then there will never really be such a thing as a missed detection. You only know if you missed a detection if you know the (discrete) times when you looked and found nothing.       

What I understand from your answer is that the PDA and JPDA is not robust against PD. 

As I know powerful target tracking algorithms like PDA/JPDA and MHT use Probability of Detection(PD) in their association computation. And these methods have been successfully applied for several practical applications. 

Do these practical applications use for example the JPDA approach with some modifications to it to increase robustness with respect to PD? 

Are there any methods that doesn't rely on PD?

I would say that all of those methods are fairly robust with respect to PD, i.e. they are not overly sensitive to it. And there are simple ways of estimating it. See ...  

Article Powerful Test Statistic for Track Management in Clutter

Thanks for your great consultant. 

The focus of your paper is on tuning the knowledge about the probability of detection.  So the uncertainty in PD is an important issue in data association problem and you solved it by estimating the probability of detection. In other word, you improved robustness of MHT with respect to PD by estimating PD.

Hello. The probability of detection does not have to be used explicitly in target tracking, as in the simple nearest neighbour Kalman filter. As soon as you wish to provide a probabilistic model of the measurement process, as used by a PDAF, for instance, you have to use PD and PFA or equivalent. The question is to what extent this simple and IID (independent/identically distributed at each time) model is valid. You could potentially model data loss as a detection probability but it is clearly correlated in time, i.e. once you lose data you are more likely to keep losing it. I would think a hybrid system would be a better model. Have a look at the link below (applied to manoeuvring target tracking, but the system model should be relevant to you).

Technical Report A Survey of Manoeuvring Target Tracking Methods

Thanks Dear Graham

Do you suggest that considering the effect of data loss (due to sensor fault or network load) separately from PD is of practical importance? 

Are there any papers that consider data loss (due to sensor fault or network load) and PD simultaneously in their trackers design?

Hello. Yes - it sounds like an important problem in certain applications. I am sure there are lots of references on this type of problem, e.g., for array processing (beamforming) and for multi-sensor tracking. Google is your friend :-)

Hello, I have googled a lot but I couldn't find any paper that exactly models 

1- Data loss (due to network)

and

2- Data loss (probability of detection), 

as a separate entity.

Please give me a reference to bright me.

Thanks.

Hello. I would think that most of the research in the last 10 years or so on "out-of-sequence" measurements is effectively about data loss or at least delays in multi-sensor networks. OTOH, if you can't find anything directly relevant, maybe you have a good idea to pursue for your own research :-)

Hello, 

In multi-sensor target tracking systems, measurements from the same target can arrive out of sequence and there are some good methods to handle this situation. But when we have only a single sensor (camera sensor) to track multiple targets,  and due to some problems (other than PD),  our measurements lost, How can we handle this situation?

Hello. The date loss and the PD are different concept. The PD is based on signal detection. But date loss is caused by sensors fail to detecte target, or the date missing during transmission. Furthermore,  date loss means the received date is a random value. But it is different from the true target is not detected, even the target is not detected, the signal received by sensors is noise which is located in the range of sensor 's detection. 

Therefore, I think date loss is different from PD. And PD is uesed in signal processing, also it can used in date processing. But date loss can only used in date processing.  Even in date processing,  the two parameters are independent and not the same.

"I couldn't find any paper that exactly models 1- Data loss (due to network) and 2- Data loss (probability of detection)"

I think Yongjian provided a good answer. Paraphrasing: data loss is not the same as failure to detect. In the "good old days" of radar, when detection theory was developed to model the achievable performance of a surveillance system, there was no such thing as "data loss", since it was all continuous-time systems (analogue electronics) and there was signal and noise.

Data is inherently a digital concept. The idea of data loss, in my experience, means that the data that you would normally expect to get at the receiver are not available. This can be due to failure of a component in the receiver or, in if there is a network, due to data loss in the network.

I am not aware of modelling of data loss due to component failure in a receiver as it impacts tracking. However, data loss, i.e., packet loss, in a digital communication network is a well studied problem. You will find many references in IEEE Comms for instance, mostly to do with buffer overflows and queuing theory.

The combination of data loss and tracking is not well studied, IMO, but I haven't looked around. The problem is that you just cannot track if data is lost, hence this is not very interesting. On the other hand, if the loss is just temporary, you may be able to post-process the resulting track fragments and joint them up. This again is only a partially studied problem.

Please upvote if you like this answer!