Generic object detection on fixed surveillance video
Systems and methods are disclosed for computer vision and object detection by extracting tracks of moving objects on a set of video sequences; selecting a subset of tracks for training; rendering a composite of each selected track into a single image; labeling tracks using the rendered images; training a track classifier by supervised machine learning using the labeled tracks; applying the trained track classifier to the remainder of the tracks; and selecting tracks classified with a low confidence by the classifier.
1. A computer vision system, comprising:
automatically extracting tracks of moving objects on a set of video sequences;
choosing a video frame belonging to the track;
overlaying onto said video frame, the pixels of selected bounding boxes of a tracked object at key positions along the tracks; and
highlighting a bounding box of the overlaid objects;
selecting a subset of tracks for training;
rendering each training track into a composite image;
labeling each track[s] using the rendered composite image[s] in one shot with an operator;
training a track classifier by supervised machine learning using the labeled tracks;
applying the trained track classifier to the remainder of the tracks; and
selecting tracks classified with a low confidence by the classifier for a subsequent labeling; and
efficiently labeling video data for training purposes and minimizing operator intervention.
2. The system of claim 1 , comprising performing background subtraction to extract the moving objects on each video frame and performing multiple hypotheses tracking to track movement of the moving objects across video frames.
3. The system of claim 1 , where the video frame is the one from the middle of the track.
4. The system of claim 1 , where the moving objects are selected from the track without overlap between overlaid objects.
5. The system of claim 1 , comprising using a graphical user interface (GUI) that displays the rendered composite image of a track and using buttons and keyboard shortcuts for labeling of the track and advancing to the next track to label with a single mouse-click or keystroke.
6. The system of claim 5 where the GUI further comprises a button or keyboard shortcut to go back to the previous track to correct for a labeling mistake.
7. The system of claim 1 , where the track classifier comprises an object classifier trained with the object's instance of the tracks, and a track classifier that uses as input features computed in part from the outputs of the object classifier at each frame of the track.
8. The system of claim 7 , where the object classifier takes as input multi-scale histogram of Gabor features from the pixels of the bounding boxes of objects.
9. The system of claim 8 , where the object classifier is a Support Vector Machine (SVM).
10. The system of claim 8 , where the object classifier is a Multi-Layer Perceptron (MLP) with one output per each class of objects.
11. The system of claim 10 , wherein the track classifier takes as input feature histograms of the Gaussian-smoothed time-series of each MLP output, and produces one output per each class of object.
12. The system of claim 11 , where the track MLP further takes as input geometric features of the track, such as statistics of size, motion and deformation.
13. The system of claim 7 , where the object classifier comprises a Convolutional Neural Network (CNN) taking as input pixels of bounding boxes of objects.
14. The system of claim 1 , comprising selecting low-confidence tracks by:
converting the outputs of the classifier to probabilities; and
selecting tracks for which no output has a high probability.
15. The system of claim 1 , where the objective is that no classification produces class probabilities below a threshold.
16. The system of claim 1 , where the objective is a given performance level on a validation set.
17. The system of claim 1 , where the objective is that a certain percentage of the tracks have been labeled.
18. The system of claim 1 , comprising capturing video sequences from fixed cameras.
19. An assisting system of a labeling task, comprising:
extracting tracks of moving objects on a set of video sequences, the set of video sequences being taken by a fixed camera, the track being movement of a moving object across a field of view of the fixed camera;
choosing a video frame belonging to the track;
overlaying onto said video frame, the pixels of selected bounding boxes of a tracked object at key positions along the tracks;
highlighting a bounding box of the overlaid objects;
rendering each extracted track into a single composite image, the single composite image representing an entire track;
accepting an operation for the composite image;
labeling the entire track represented by the single composite image in response to the operation; and
providing training data to a training classifier on the basis of the labeled tracks; and
efficiently labeling video data for training purposes and minimizing operator intervention.
20. The system of claim 19 , comprising performing background subtraction to extract the moving objects on each video frame.
21. The system of claim 19 , where the operation is a single action.
22. The system of claim 19 , where the composite image contains a plurality of bounding boxes, each bounding box representing the moving object.
23. The system of claim 19 , where the moving objects are selected from the track without overlap between overlaid objects.