All-weather thermal-image pedestrian detection method

Patent metadata

Field	Value
Publication number	US-10198657-B2
Application number	US-201615375438-A
Country	US
Kind code	B2
Filing date	Dec 12, 2016
Priority date	Dec 12, 2016
Publication date	Feb 5, 2019
Grant date	Feb 5, 2019

Abstract

An all-weather thermal-image pedestrian detection method includes (a) capturing diurnal thermal images and nocturnal thermal images of a same pedestrian and non-pedestrian object in a same defined block to create a sample database of thermal images, wherein the sample database comprises pedestrian samples and non-pedestrian samples; (b) performing LBP encoding on the pedestrian samples and the non-pedestrian samples, wherein complementary LBP codes in the same defined block are treated as identical LBP codes; (c) expressing the LBP codes in the same defined block as features by a gradient direction histogram (HOG) to obtain feature training samples of the pedestrian samples and the non-pedestrian samples; (d) entering the feature training samples into a SVM to undergo training by Adaboost so as to form a strong classifier; and (e) effectuating pedestrian detection by searching the strong classifiers in thermal images with sliding window technique to detect for presence of pedestrians.

First claim

What is claimed is:

1. An all-weather thermal-image pedestrian detection method, comprising the steps of:
(a) capturing diurnal thermal images and nocturnal thermal images of a same pedestrian and non-pedestrian object in a same defined block to create a sample database of thermal images, wherein the sample database comprises a plurality of pedestrian samples and a plurality of non-pedestrian samples;
(b) performing local binary patterns (LBP) encoding, in the same defined block, on the pedestrian samples and the non-pedestrian samples in the sample database, wherein complementary LBP codes in the same defined block are treated as identical LBP codes;
(c) expressing the LBP codes in the same defined block as features by a gradient direction histogram (HOG) to obtain feature training samples of the pedestrian samples and the non-pedestrian samples;
(d) entering the feature training samples into a support vector machine (SVM) to undergo training by adaptive boosting (Adaboost) so as to form a strong classifier, wherein step (d) comprises the steps of:
(d1) scanning a plurality of defined regions, of different sizes, on the whole image;
(d2) allowing each defined block to obtain feature training samples of a plurality of pedestrian samples and non-pedestrian samples by steps (a)˜(c);
(d3) entering the feature training samples into the SVM to undergo training so as to obtain a plurality of weak classifiers; and
(d4) searching, by Adaboost computation, the weak classifiers for at least a strong classifier with key positions of pedestrians; and

(e) effectuating pedestrian detection by searching the strong classifiers in thermal images with sliding window technique to detect for presence of pedestrians.

2. The all-weather thermal-image pedestrian detection method of claim 1 , wherein step (b) comprises the sub-steps of:
(b1) performing LBP encoding on diurnal thermal images and nocturnal thermal images of the pedestrian samples and the non-pedestrian samples; and
(b2) treating complementary LBP codes in the same defined block as identical LBP codes.

3. The all-weather thermal-image pedestrian detection method of claim 1 , wherein step (c) comprises the sub-steps of:
(c1) dividing the same defined block into a plurality of block regions;
(c2) dividing each block region into a plurality of unit regions, wherein the unit regions each have a plurality of LBP codes;
(c3) calculating gradient intensity and gradient direction of all the LBP codes in each block region; and
(c4) performing vote counting on all the LBP codes in each unit region according to their gradient intensity and gradient direction to obtain the feature vector of each unit region, wherein the feature vectors of the unit regions together form HOG features of the block regions, respectively, and the HOG features of the block regions form the HOG features of the same defined block to therefore obtain features training samples of the pedestrian samples and the non-pedestrian samples.

4. The all-weather thermal-image pedestrian detection method of claim 1 , wherein step (e) comprises the sub-steps of:
(e1) scanning the strong classifiers in the thermal images with sliding window technique;
(e2) treating the blocks of the strong classifier as LBP codes;
(e3) expressing the LBP codes as features by HOG; and
(e4) entering the HOG features into the SVM classifier to undergo pedestrian recognition.

Assignees

• Nat Chung Shan Inst Science & Tech

Inventors

• Huang Shih-Shinh
• Chien Shih-Che
• Chang Feng-Chia
• Hsiao Chien-Hao
• Hsiao Yu-Sung

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/50

  by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis · CPC title

• H04N19/103 Primary

  Selection of coding mode or of prediction mode · CPC title

• G06F18/2411

  based on the proximity to a decision surface, e.g. support vector machines · CPC title

• H04N19/176

  the region being a block, e.g. a macroblock · CPC title