Issues for code generation of trackerPHD and trackingSe​nsorConfig​uration using Matlab Coder

Hello,

I was trying to generate C++ for the following function codegenTracker. After a careful study of the extended capabilities section of the trackerPHD document and a lot of experiments, I was able to generate the code successfully.

However, I found that I have to set the number of sensors, which is n in the code below expectily ( to 3, in mycase), so that it does not give me the error: 'Loop unrolling failed because the number of iterations is unknown' during 'Check for Run-Time Issues'. From an application point of view, I prefer n to be specified in a configuration file, so that I could load n like the tempConfig struct in the code below, instead of set it to a number expectily. For example, if I want to change the ClutterDensity property of trackingSensorConfiguration, I can just change it in my configuration file without the need to change it in my matlab code and regenerate the entire code.

Similarly, for trackerPHD, I have to specify the properties expectily. If I try to set a property of trackerPHD,say, BirthRate, through the configuration file, I got the following error: 'Failed to compute constant value for nontunable property 'BirthRate'. In code generation, nontunable properties can only be assigned constant values'.

How do I modify my code so that I do not need to specify all the parameters expectily and it still supports code generation? Thanks in advance. I have attach my code for code generation, which is based on the matlab project 'Asynchronous Angle-only Tracking with GM-PHD Tracker'.

function tracks = codegenTracker(detections, configs, time) %#codegen

tracks = repmat(DummyTracks(),0,1);
coder.varsize('tracks',[Inf,1],[1,0]);

persistent tracker;
if isempty(tracker)
    
    % read parameters from the configuration file, tempConfig is a struct that
    % contains these parameters.
    
    tempConfig = config.readSensorConfig();
    % Prefer Load n from a configuration file instead of set it to a number
    % expectily
    % 
    n = 3;
    sensorConfigs = cell(n,1) ;                  
    for i = 1 : n
        sensorConfigs{i,1} = trackingSensorConfiguration("SensorIndex",single(i), ...
            "IsValidTime",true, ...
            "SensorLimits", tempConfig.SensorLimits,...
            "SensorResolution", tempConfig.SensorResolutions,...
            "SensorTransformFcn",'cvmeas', ...
            "SensorTransformParameters", DummyMps(),...
            "ClutterDensity", tempConfig.ClutterDensity, ...
            "MinDetectionProbability", tempConfig.MinDetectionProbability,...
            "FilterInitializationFcn", 'initcvAngleOnlyGMPHD',...
            "MaxNumDetsPerObject", uint32(1), ...
            "MaxNumDetections", uint32(100));
    end
    
    tracker = trackerPHD(SensorConfigurations = sensorConfigs, ...
        BirthRate = 0.1, ...
        DeathRate = 1e-6, ...
        AssignmentThreshold = 25 , ...
        ExtractionThreshold = 0.85, ...
        ConfirmationThreshold = 0.95, ...
        DeletionThreshold =  1e-3, ...
        MergingThreshold = 25, ...
        LabelingThresholds = [1.1 1 0.8], ...
        HasSensorConfigurationsInput = true, ...
        MaxNumSensors = 20, ...    
        MaxNumTracks = uint32(1000), ... 
        MaxNumComponents = uint32(5000));  
end

if tracker.isLocked() || ~isempty(detections)
    tracks = tracker(detections, configs, double(time));
end

end

Loop unrolling failed because the number of iterations is unknown.

Answer 1

[Prashant Arora]

Hi Lechi,

The tracker defines many of these properties as "nontunable", which means that you cannot change them after C/C++ code is generated. You will need to regenerate code with new set of values.

There are some properties which are possible to make "tunable". For example, BirthRate, DeathRate etc, but they are not as of R2023b. However, some properties control the size, datatype and memory allocation of datatypes in C/C++. For example, it will be difficult to support "MaxNumTracks", "MaxNumSensors" etc. as tunable properties. I will create an enhancement request to support changing some of these properties after code generation.

Regarding SensorConfigurations:

The number and properties of the sensor configurations can be tuned on the fly in a way. You can pre-define the maximum sensors as Nmax. You can construct the tracker with Nmax configurations. However, only n out of those nMax sensors can be active during a run. To do this, you can use configs input to the tracker. By setting the IsValidTime of the sensors "inactive" during this run as false, you can indicate that these trackers only n sensors are being used during a run.

Hope this helps.

Thanks,

Prashant

[Lechi Li]

Thank you very much !!!

[Alan Bindemann]

I am trying to code-generate a trackerPHD tracker using the generateCode function under 2024b. However the build fails with the error:

Found unsupported class for variable using "coder.load" Mixed Field types in structure arrays are not supported for code generation. Type at s.detection(1).Frame must match type of s.detection(2).Frame.

When I examine the file detection.mat that was autogenerated using the generateCode call, it only has a scalar detection field (i.e. s.detection(2) does not exist).

The code that creates the detection.mat file is behind p-code, so I can't determine the issue.