Awesome

Overview

A multi-pedestrian tracker as proposed in the paper:

Führ, G. ; Jung, C. R. Combining Patch Matching and Detection for Robust Pedestrian Tracking in Monocular Calibrated Cameras. Pattern Recognition Letters (Special Issue). 2013.

If you use the following code in our experiments, please cite the above publication.

Videos of the tracker performance and information about it can be found in: http://inf.ufrgs.br/~gfuhr/?file=research Soon, there will be extensions of the code and improvements in this webpage.

For question about the code/method please contact Gustavo Führ at gfuhr@inf.ufrgs.br.

Requirements

The source code was tested in MATLAB 64-bin in version 7.12 (R2010a) in Linux x64. However, you should not have problems in running the code in different platforms and newer versions of MATLAB. This applications includes the person detector developed by Piotr Dóllar, which can be found in his personal website.

Running the code

1. Unpack the code
2. Run the main.m file using MATLAB

All the important configurations of the tracker are set in the main.m. This package contains an example main.m with configurations used in the datasets presented in the paper. Please not that you should download the sequences separately.

The folder calibration/ provides the calibration files for the PETS and TownCentre dataset used in the paper.

Configuration of main.m

• options.init_n_patches The number of patches in which a pedestrian is segmented. (default: 6)

• options.init_aspect_ratio This is the aspect_ratio used to compute the width of the patches at initialization. (default: 0.26)

• options.init_max_target_overlap Maximum allowed ratio of overlap between different detected pedestrians -- if the overlap is above this value the detections are rejected. (default: 0.01)

• options.init_min_perc_foreground In order to avoid false positives of the person detector, only detections with significant amount of background are initialized. This value is the minimum amount of background required to consider the detection bounding box. (default: 0.4)

• options.max_world_height Maximum height in meters, for a pedestrian, in real-world coordinates. If the height of the detected pedestrian is above, the initialization is aborted. This value is used to reject false positives generated by the person detector. (default: 2.5)

• options.min_world_height Minimum height in meters, for a pedestrian in real-world coordinates. (default: 1.5)

• options.term_border_temporal_window_size Number of frames in which a target that has been tracked poorly is terminated if the target is (currently) on the borders of the image. (default: 4)

• options.term_center_temporal_window_size Number of frames in which a target that has been tracked poorly is terminated if the target is (currently) on the center of the image. (default: 16)

• options.term_no_detections_n_frames Number of frames after which the track is terminated if no detection was associated with the target (Default: 16).

• options.n_bins Number of bins used to compute the color histograms. (default: 128)

• options.detection_search_region Radius (in meters) of the search region around the target to associate pedestrians detections. (default: 1.5)

• options.s_temporal_window Number of past frames used to compute the weight of the detection vector in the WVMF. (default: 50)

• options.wvmf_gamma Gamma parameter value for the Weighted Vector Median Filter. See paper for details. (default: 0.25)

• options.motion_prior_alpha Alpha parameter to compute the motion prediction. See paper for details (default: 0.08)

• options.world_search_sample_step Step size (in pixels) that is used to sample the world candidate region after its projection. (default: 1)

• options.walking_speed Represents the maximum pedestrian walking speed expected in the sequence in m/s. (default: 1.5)

• options.relaxation_parameter Relaxation parameter to increase the size of the search region (usually used to recover from failure). (default: 0.5)

• options.show_frames Boolean field to show the results of tracking. (default: true)

• options.plot_bboxes If true, makes it so that bounding boxes are plotted during execution. (default: false)

• options.plot_patches Boolean field that determines whether or not the patches used for tracking are displayed. (default: true)

• options.plot_trajectories Variable to decide if pedestrian trajectories are displayed. (default: false)

• options.plot_target_number Boolean field that determines whether or not the pedestrian identifier is plotted in the result frames. (default: true)

• options.plot_detections Plot the pedestrian detections with colors that indicate to which target they were associated. (default: false)

• options.plot_prediction_vector Plot the projected prediction vector for each target. (default: false)

• options.save_frames Save the frames in the options.out_path folder. It requires the package export_fig to work (export_fig can be obtained at: http://www.mathworks.com/matlabcentral/fileexchange/23629-exportfig) (default: false)

• options.out_filename Character string that represents the name of the output file.

• options.out_path Path where output files should be stored.

• options.image_pref The prefix of the string used to create the filenames.

• options.file_ext The file extension of the image files.

• options.d_mask The mask of integers used to create the numeric part of the filename. For instance, if the options.image_pref = 'data/fr', options.file_ext = 'png' and options.d_mask = 3, the filenames will be created as 'data/fr001.png', 'data/fr002.png', etc.

• options.calib_filename The path to the xml file containing the camera calibration. The xml is expected to be in the format provided with the PETS dataset. The calibration/pets.xml in this package is an example of such calibration file.

• options.world_unit The measurement unit of the calibration file. The possible values are 'm' (meters), 'cm' (centimeters) and 'mm' (milimeters).

• options.video_fps The frames per second of the sequence. This is used when defining the search region in the world. See the papers for details.

• options.begin_frame The first frame that will be used in the sequence.

• options.end_frame The number of the last frame to be processed.

• options.bg_model_filename The background model file previously stored. Use this file to save the background model once it is computed. You can examples of this in the folder bg_models/

• options.init_bg_training If the background model file is not provided, then it is going to be computed at initialization. The frame range used begins in init_bg_training and ends at begin_frame.

• options.begin_frame The first frame that will be used in the sequence.

• options.end_frame The number of the last frame to be processed.

Output

A file with the trackers results (positions) is stored in a .mat file named with the parameter options.out_filename.

Enjoy and please send us feedback!