Research On Key Technologies Of Image-based Pedestrian Detection

Pedestrian detection is one of the most important topics in computer vision, which is widely applied in vehicle assistance driving and other fields.Pedestrians have the characteristics of both the rigid and soft object. They are very close to some rigid objects like cars and in some degree of deformation at the same time. Besides, pedes-trians are usually under partial occlusion. All of these problems make the pedestrian detection a challenging problem.Pedetrian detection must also consider the restraint on detecting speed,making the problem more challenge.Most of the detecting methods could be classified into two categories. The first one is based on parts. These methods usually first segment the human body into parts and train a classifier for each, and then integrate them with a certain algorithm.The main disadvantage is that the part detector itself is not so discriminative and is easily affected by occlusions and deformations.The other kind of method is based on the whole body which focus more on the features. The main disadvantage is that features used in these methods are usually manually designed and of lower level.Existing methods have achieved some result,however,there still exist areas that could be improved,specially the algorithm of integrating the part detectors in parts based models and feature designing in whole body based models. Therefore,this paper carries out researches on these areas, containing the following three aspects:Firstly, we propose a probabilistic model based on conditional random field, which is applied to integrate part detectors and supply them with extra contextual information from the overlapped parts.Therefore the detection result of a part detector is decided not only by the information of the part,but also by the detecting result of related parts, which makes the detecting result of the part detectors more reliable, thus improving the whole detecting effect.Secondly,most of the methods based on the whole body adopt manually designed features of low level, which could not represent higher lever information or making full use of the mount of images. To solve this problem, we use the sparse feature learning algorithm to compute sparse channel features and extending them with some lower level channel features, making them owning the characteristics of both low level and higher level features,while taking the detecting speed into consideration.Lastly,we takes the specificality of pedestrian detection into consideration. Pedes-trian detection is something like a two category classification and we care more about pedestrian side,at the same time a part of a pedestrian is usually restricted in a certain area.We try to learn a dictionary with a supervised method,and combine models based on parts with models based on whole body, treating the scores from a part detector as a higher level features,which improves the detecting result in a further step.