Traffic Sign Recognition: Classical HOG + SVM vs Convolutional Neural Network

Driver-assistance systems must recognise traffic signs under blur, poor light and partial occlusion. On the real German Traffic Sign Recognition Benchmark (GTSRB: 43 classes, 39,209 training images in 1,307 tracks, 12,630 official test images) the project compares the classical image-processing chain - ROI crop, 32 x 32 resize, histogram of oriented gradients (HOG, OpenCV) with an optional HSV colour histogram, linear or RBF-kernel support vector machine - with a small convolutional neural network trained on a laptop CPU. Every choice is made on track-grouped validation data and the official test set is used once: CNN 96.44 % (deployed), HOG + colour + linear SVM 96.15 %, HOG + linear SVM 95.72 %, HOG + RBF SVM 95.23 %. A random split would overstate the accuracy (97.61 % vs 95.33 % in cross-validation) because frames of the same physical sign land on both sides. A robustness sweep shows the CNN degrading less under blur, noise and rotation, while HOG is almost immune to darkness. GNU Octave reproduces the HOG + linear SVM chain; a complete MATLAB version (written, not run) and a web application with five databases are included.

Category: Electronics and Communication Engineering (Image Processing and Computer Vision, Histogram of Oriented Gradients, Support Vector Machines, Convolutional Neural Networks, Driver-Assistance Systems, Robustness Testing, OpenCV, GNU Octave, MATLAB, Python)

Electronics & Communication Engineering Project E10 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.