| 250 | // ---------------------------------------------------------------------------------------- |
| 251 | |
| 252 | int main() |
| 253 | { |
| 254 | initialize_object_properties(); |
| 255 | |
| 256 | |
| 257 | // Get some training data. Here we sample 5 independent track histories. In a real |
| 258 | // world problem you would get this kind of data by, for example, collecting data from |
| 259 | // your sensor on 5 separate days where you did an independent collection each day. |
| 260 | // You can train a model with just one track history but the more you have the better. |
| 261 | std::vector<track_history> data; |
| 262 | data.push_back(make_random_tracking_data_for_training()); |
| 263 | data.push_back(make_random_tracking_data_for_training()); |
| 264 | data.push_back(make_random_tracking_data_for_training()); |
| 265 | data.push_back(make_random_tracking_data_for_training()); |
| 266 | data.push_back(make_random_tracking_data_for_training()); |
| 267 | |
| 268 | |
| 269 | structural_track_association_trainer trainer; |
| 270 | // Note that the machine learning tools have a parameter. This is the usual SVM C |
| 271 | // parameter that controls the trade-off between trying to fit the training data or |
| 272 | // producing a "simpler" solution. You need to try a few different values of this |
| 273 | // parameter to find out what setting works best for your problem (try values in the |
| 274 | // range 0.001 to 1000000). |
| 275 | trainer.set_c(100); |
| 276 | // Now do the training. |
| 277 | track_association_function<detection> assoc = trainer.train(data); |
| 278 | |
| 279 | // We can test the accuracy of the learned association function on some track history |
| 280 | // data. Here we test it on the data we trained on. It outputs a single number that |
| 281 | // measures the fraction of detections which were correctly associated to their tracks. |
| 282 | // So a value of 1 indicates perfect tracking and a value of 0 indicates totally wrong |
| 283 | // tracking. |
| 284 | cout << "Association accuracy on training data: "<< test_track_association_function(assoc, data) << endl; |
| 285 | // It's very important to test the output of a machine learning method on data it |
| 286 | // wasn't trained on. You can do that by calling test_track_association_function() on |
| 287 | // held out data. You can also use cross-validation like so: |
| 288 | cout << "Association accuracy from 5-fold CV: "<< cross_validate_track_association_trainer(trainer, data, 5) << endl; |
| 289 | // Unsurprisingly, the testing functions show that the assoc function we learned |
| 290 | // perfectly associates all detections to tracks in this easy data. |
| 291 | |
| 292 | |
| 293 | |
| 294 | |
| 295 | // OK. So how do you use this assoc thing? Let's use it to do some tracking! |
| 296 | |
| 297 | // tracks contains all our current tracks. Initially it is empty. |
| 298 | std::vector<track> tracks; |
| 299 | cout << "number of tracks: "<< tracks.size() << endl; |
| 300 | |
| 301 | // Sample detections from 3 objects. |
| 302 | std::vector<detection> dets = make_random_detections(3); |
| 303 | // Calling assoc(), the function we just learned, performs the detection to track |
| 304 | // association. It will also call each track's update_track() function with the |
| 305 | // associated detection. For tracks that don't get a detection, it calls |
| 306 | // propagate_track(). |
| 307 | assoc(tracks, dets); |
| 308 | // Now there are 3 things in tracks. |
| 309 | cout << "number of tracks: "<< tracks.size() << endl; |
nothing calls this directly
no test coverage detected