Real AdaBoost HOG 1 1 1, 2 1 Real AdaBoost HOG HOG Real AdaBoost HOG A Method for Reducing number of HOG Features based on Real AdaBoost Chika Matsushima, 1 Yuji Yamauchi, 1 Takayoshi Yamashita 1, 2 and Hironobu Fujiyoshi 1 We propose an efficient method for reducing the number of HOG features based on Real AdaBoost. The proposed method can reduce the amount of memory required by converting the HOG features used in human detection to a binary pattern. Converting to a binary pattern, however, causes the problem of a sparse probability density distribution used in classification. Nevertheless, a dense probability density distribution can be achieved by using Real AdaBoost to integrate the binary pattern during training. To confirm the effectiveness of the proposed method, we conducted evaluation experiments and compared the amounts of memory required. The results show that the detection accuracy of the HOG features is retained, but less memory is required for the processing. 1. ITS Edge of Orientation Histograms (EOH) 1) Histograms of Oriented Gradients (HOG) 2) 2) 7) Large Scale Integration (LSI) Field Programmable Gate Array (FPGA) 8) LBG Real AdaBoost 4),9) HOG HOG HOG Real AdaBoost 2 Real AdaBoost HOG 3 4 HOG 1 Chubu University 2 OMRON Corporation 1 c 2009 Information Processing Society of Japan
{ f x(x, y) =L(x +1,y) L(x 1,y) f y(x, y) =L(x, y +1) L(x, y 1) Fig. 1 1 Binarization of HOG features by fixed value. 2. Real AdaBoost HOG Real AdaBoost HOG HOG HOG Real AdaBoost 2.1 1 HOG HOG 2.1.1 HOG Histograms of oriented gradients(hog) 2) HOG L m θ m(x, y) = f x(x, y) 2 + f y(x, y) 2 (1) θ(x, y) =tan 1 f y(x, y) f x(x, y) (2) m θ 0 360 0 180 8 (3) v v = (ɛ =1) (3) k v 2 i + ɛ2 i=0 v HOG k HOG 2.1.2 HOG HOG 30 60 5 5 3 3 2,880 1 8 ( ) 23,040 1 (720 480 ) 50,000 1.07 GB FPGA LSI 2.1.3 HOG 2.1.1 HOG 1 1 8 HOG { 1 v d th b d = (4) 0 otherwise b d(1 8) th HOG 30 60 5 5 3 3 360 1 1 ( ) HOG 64 2 c 2009 Information Processing Society of Japan
Fig. 3 3 1 An example of probability density distribution at 1st round training. 2 Real AdaBoost Fig. 2 Algorithm of reducing number of feature based on Real AdaBoost. 1 th 0.03 2.2 Real AdaBoost Real AdaBoost 2(a) 2(a) 1. N 2(a) 2. 2(a) 3.1. ( ) ( ) W + W W + W 1 D t 3(a) W + W W + W h(x) ɛ 0 9) ɛ =1/N 2(a) 3.4. W + W Z m Z m 2(a) 4. Z t 2(a) 5. H(x) 2.2.1 HOG Real AdaBoost 1 W + W 3(a) W + W HOG 0 h(x) h(x) = 1 ( W+ + ɛ ln ɛ = 1 ) (5) 2 W + ɛ N H(x) h(x) 3 c 2009 Information Processing Society of Japan
Fig. 4 4 Example of searching for integration pattern. 2.2.2 Real AdaBoost 2.2.1 2(b) 3(b) W + W 0 0 2(b) 1. j λ 2.0 j 2(b) 3. j 4(b) 1 8 Fig. 5 5 Examples of output of strong classifier. j j j r 2(b) 4. r ĵ j ĵ Real AdaBoost 3(b) 5 6 Fig. 6 Example of database. H(x) 3. 4 c 2009 Information Processing Society of Japan
1 1 Table 1 Classification rate and memory size for processing an image. ( ) [%] [MB] HOG 2) 75.6 1,098.63 70.8 198.75 65.8 128 76.5 64 76.5 17.17 32 74.2 Fig. 7 7 DET DET curves of evaluation experiments. 3.1 6 2,054 6,258 1,024 1,234 3.2 HOG 400 Detection Error Tradeoff(DET) DET 1 (720 480 ) 50,000 3.3 7 DET 1 5.0% 2 n (n =5, 6, 7) 7 HOG 1 64 HOG 98.0% 8 Fig. 8 Visualization examples of selected feature by training. 3.4 Real AdaBoost HOG HOG 8 8(a) HOG 8(b) HOG 5 c 2009 Information Processing Society of Japan
HOG 4. HOG Real AdaBoost 2 10) HOG 4.1 2.1.3 HOG 9 8 HOG (6) { 1 v d v d b d = (6) 0 otherwise 9 HOG Fig. 9 Binary pattern using HOG feature of different cells. b d v HOG 9 HOG HOG 10 2 HOG 9 HOG HOG 2 HOG 2.2.2 Real AdaBoost 4.2 HOG 10 Fig. 10 Binary pattern including bit shift of orientation. 4.2.1 HOG HOG 6 c 2009 Information Processing Society of Japan
Fig. 11 11 DET DET curves of evaluation experiments by binarization of HOG features. 3 64 400 DET 4.2.2 11 DET 2 5.0% 11 HOG HOG HOG 2 5.0% 76.5% HOG ( ) 82.0% 6.4% 4.2.3 12 12(b) 2 HOG 12(c) 12(b) 12(b) (c) 400 12(b) 4.0% 12(c) 12.0% HOG 12 Fig. 12 Visualization example of selected binary pattern by training. 7 c 2009 Information Processing Society of Japan
Table 2 2 Classification rate for each method. [%] 76.5 ( ) 68.3 76.5 74.9 82.0 5. Real AdaBoost HOG HOG HOG HOG Real AdaBoost 98.0% HOG 6.4% Single Image by Bayesian Combination of Edgelet Part Detectors, IEEE International Conference on Computer Vision, pp.90 97 (2005). 5) Mu, Y., Yan, S., Liu, Y., Huang, T. and Zhou, B.: Discriminative local binary patterns for human detection in personal album, IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.1 8 (2008). 6) Munder, S. and Gavrila, D.M.: An Experimental Study on Pedestrian Classification, IEEE Transactions on Pattern Analysis and Machine Intelligence, pp.1863 1868 (2006). 7) Hou, C., Ai, H. and Lao, S.: Multiview Pedestrian Detection Based on Vector Boosting, Eighth Asian Conference on Computer Vision, pp.220 229 (2007). 8) Linde, Y., Buzo, A. and Gray, R.: An Algorithm for Vector Quantizer Design, IEEE Transactions on Communications, Vol.28, No.1, pp.84 95 (1980). 9) Schapire, R.E. and Singer, Y.: Improved Boosting Algorithms Using Confidencerated Predictions, Machine Learning, pp.297 336 (1999). 10) Joint HOG 2 AdaBoost 14 SSII08 (2008). 1) Levi, K. and Weiss, Y.: Learning Object Detection from a Small Number of Examples: the Importance of Good Features, IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.53 60 (2004). 2) Dalal, N. and Triggs, B.: Histograms of Oriented Gradients for Human Detection, IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Vol.1, pp.886 893 (2005). 3) Chen, Y.T. and Chen, C.S.: Asian Cascade of Feed-Forward Classifiers for Fast Pedestrian Detection, Eighth Asian Conference on Computer Vision, pp.905 914 (2007). 4) Wu, B. and Nevatia, R.: Detection of Multiple, Partially Occluded Humans in a 8 c 2009 Information Processing Society of Japan