Detecção de pedestres utilizando descritores de orientação do gradiente e auto similaridade de cor

Abstract

Pedestrian detection is a key problem in our days, having a large number of applications with potential to make better the quality of life of our society. Some of these applications can be found in driver assistance systems, people recognition in photos and videos, and surveillance. Nowadays, there is a large number of researches in this area, generating a lot of ramifications in the state of the art for pedestrian detection. This dissertation presents a pedestrian detection system in non-controlled environments based on sliding windows. Systems of this type are based on two major blocks: one for feature extraction and other for window classification. Two techniques for feature extraction are used: HOG (Histogram of Oriented Gradient) and CSS (Color Self Similarities), and to classify windows we use linear SVM (Support Vector Machines). Beyond these techniques, we use: mean shift and hierarchical clustering, to fuse multiple overlapping detections; and bilateral filter, to preprocess the image. The results obtained bu testing the dataset INRIA Person show that the proposed system, using only HOG descriptors, achieves better results over similar systems, with a log average miss rate equal to 41.8%, against 46% of the literature. These results were possible due to the cutting of the final detections to better adapt them to the modified annotations, and some modifications on the parameters of the descriptors. The addition of the modified CSS descriptor to the HOG descriptor increases the efficiency of the system, leading to a log average miss rate equal to 36.2%, when classifying each descriptor separately.