Vision-based system for pedestrian recognition using a tuned SVM classifier

Pedestrian recognition is one of the main advantages of the currently introduced autonomous cars. It is expected that millions of lives will be saved just by implementing this technology in real roads. We study this problem from two points of view, i.e., the recognition algorithm and the data. A trained binary classifier based on a tuned RBF-kernel SVM is used for predicting pedestrians on new scenarios. It is shown that tuning this classifier improves significantly the performance when compared with an SVM with linear kernel. The images are pre-processed using the HOG algorithm in order to get a pedestrian descriptor. The prediction is evaluated using the F₁ score instead of the accuracy, because it presents a better estimation of the model performance, and yields a better way of tuning the model. The model is validated using the cross-validation method; the averaged accuracy and F₁ score reached were 95% and 96.3% respectively. A database composed of 5,000 pedestrian and non-pedestrian images is used for training and testing the classifier. Several pedestrian images are analyzed after applying the algorithm to determine how the database should be completed in order to improve the detection in actual road scenarios.