Traffic sign recognition using local directional histogram of oriented gradients

Histogram of Oriented Gradients (HOG) describes an image gradient by calculating vertical and horizontal gradient magnitudes and directions. HOG uses a one-dimensional (1D) centered derivative mask [−1, 0, +1] for horizontal gradient and its rotations at 90o for vertical gradient. This technique only considers four neighboring pixels while calculating image gradient at a particular pixel. Every pixel in an image carries subtle information and therefore all pixels should be considered when deriving image gradient. Therefore, given a pixel pi, all its N = 2(d+1) neighbors should be considered when calculating the gradient at distance d from pi. This paper proposes Local Directional Histogram of Oriented Gradients (LD-HOG), which, given pixel pi, it calculates the gradient at distance d = 1 from pi by considering all the eight neighbors of pi. The proposed operator calculates the image gradient at 0o, 45o, 90o and 1350. These image gradients are used to generate two HOG histograms. Maximum pooling techniques were applied to combine the two histograms. Experimental results on the German traffic sign detection benchmark (GTSDB) dataset show that LD-HOG (average precision = 0.90, average recall = 0.90 and average F1-score =0.90) out performs HOG (average precision = 0.84, average recall = 0.82 and average F1-score = 0.83) in traffic sign recognition. The averages of the two extractors (HOG and LD-HOG) were calculated from experimental results after applying Support Vector Machine (SVM), Random Forest (RF) and Decision Tree (DT) machine learning classifiers. Stratified K-Fold Cross-Validation was done on the proposed LD-HOG using SVM, RF and DT. Validation results show that SVM performed better with 99 percent, followed by RF with 96 percent. DT was had 76 percent.