lib/ccv_tld.c

ccv_tld_new

ccv_tld_t* ccv_tld_new(ccv_dense_matrix_t* a, ccv_rect_t box, ccv_tld_param_t params)

Create a new TLD tracking instance from a given first frame image and the tracking rectangle.

• a : the first frame image.
• box : the initial tracking rectangle.
• params : a ccv_tld_param_t structure that defines various aspects of TLD tracker.

ccv_tld_param_t

TLD is complex enough that I will divide its parameters into sections:

Short-term lucas-kanade tracking parameters

• win_size : the window size to compute optical flow
• level : level of image pyramids
• min_eigen : the minimal eigenvalue for a valid optical flow computation
• min_forward_backward_error : the minimal forward backward error

Image pyramid generation parameters (for scale-invariant object detection)

• interval : how many intermediate images in between each image pyramid level (from width => width / 2)
• shift : how much steps sliding window should move

Samples generation parameters

• min_win : the minimal window size of patches for detection
• include_overlap : above this threshold, a bounding box will be positively identified as overlapping with target
• exclude_overlap : below this threshold, a bounding box will be positively identified as not overlapping with target

Ferns classifier parameters

• structs : how many ferns in the classifier
• features : how many features for each fern

Nearest neighbor classifier parameters

• validate_set : the conservative confidence score will be only computed on a subset of all positive examples, this value gives how large that subset should be
• nnc_same : above this threshold, a given patch will be identified as the same
• nnc_thres : the initial threshold for positively recognize a patch
• nnc_verify : the threshold for a tracking result from short-term tracker be verified as a positive detection
• nnc_beyond : the upper bound threshold for adaptive computed threshold
• nnc_collect : the threshold that a negative patch above this will be collected as negative example
• bad_patches : how many patches should be evaluated in initialization to collect enough negative examples

Deformation parameters to apply perspective transforms on patches for robustness

• new_deform : number of deformations should be applied at initialization
• new_deform_angle : the maximal angle for x, y and z axis rotation at initialization
• new_deform_scale : the maximal scale for the deformation at initialization
• new_deform_shift : the maximal shift for the deformation at initialization
• track_deform : number of deformations should be applied at running time
• track_deform_angle : the maximal angle for x, y and z axis rotation at running time
• track_deform_scale : the maximal scale for the deformation at running time
• track_deform_shift : the maximal shift for the deformation at running time

Speed up parameters

• top_n : only keep these much positive detections when applying ferns classifier
• rotation : when >= 1, using “rotation” technique, which, only evaluate a subset of sliding windows for each frame, but after rotation + 1 frames, every sliding window will be evaluated in one of these frames.

ccv_tld_track_object

ccv_comp_t ccv_tld_track_object(ccv_tld_t* tld, ccv_dense_matrix_t* a, ccv_dense_matrix_t* b, ccv_tld_info_t* info)

ccv doesn’t have retain / release semantics. Thus, a TLD instance cannot retain the most recent frame it tracks for future reference, you have to pass that in by yourself.

• tld : the TLD instance for continuous tracking
• a : the last frame used for tracking (ccv_tld_track_object will check signature of this against the last frame TLD instance tracked)
• b : the new frame will be tracked
• info : a ccv_tld_info_t structure that will records several aspects of current tracking

ccv_tld_info_t

• perform_track : whether we performed tracking or not this time
• perform_learn : whether we performed learning or not this time
• track_success : if we have a successful tracking (thus, short term tracker works)
• ferns_detects : how many regions passed ferns classifier
• nnc_detects : how many regions passed nearest neighbor classifier
• clustered_detects : after cluster, how many regions left
• confident_matches : how many matches we have outside of the tracking region (may cause a re-initialization of the short term tracking)
• close_matches : how many matches we have inside the tracking (may cause a new learning event)

ccv_tld_free

void ccv_tld_free(ccv_tld_t* tld)

• tld : the TLD instance to be freed.

‹ back

comments powered by Disqus