In a pedestrian detection system, the most criticalrequirement is to quickly and reliably determinewhether a candidate region contains a pedestrian.It is essential to design an effective classifier forpedestrian detection. Until now, most of theexisting pedestrian detection systems only adopt asingle and non-cascaded classifier However, sincethe scene is complex and the candidate regions aretoo many (in our experiments, there are more than 40,000 candidate regions); it is difficult to make therecognition both accurate and fast with such anon-cascaded classifier

The most critical requirement of a pedestriandetection system is to quickly recognize pedestriansin an image. However, the huge number ofcandidateregions and the complexity of scenes usually makethe recognition slow and unreliable. An efficientclassifier is neededfor a pedestrian detection system.In this paper, a decomposed SVM algorithm is usedto train a classifier for pedestrian detection. Thealgorithm is stable and suitable for training aclassifier with a large number of samples and thederived classifier is very efficient. Meanwhile,considering that our system is based on a singlecamera and the scenes are always complex, it isdifficult to train a good classifier only with shapefeatures. To solve these problems, we integrate shapeinformation with motion information to compose afeature set and use it to train a classifier.Experiments show that our system based on thisclassifier works very well. Furthermore, we discussthe effect of applying motion features. With a properpercentage, motion features will be a goodcomplement of the shape features in complex scenes.Comparison between application of shape featuresand application of both shape and motion featuresshows the advantage ofour method.

Optical camera based detection method is a popular system to fulfillpedestrian detection; however, it is difficult to be used to detect pedestrians incomplicated environment (e.g. rainy or snowy weather conditions). The difficultiesmainly include: (1) The light is much weaker than in sunny days, thereforeit is more difficult to design an efficient classification mechanism; (2)Since a pedestrian always be partly covered, only using its global features (e.g.appearance or motion) may be mis-detected; (3) The mirror images on wet roadwill cause a lot of false alarms. In this paper, based on our pervious work, weintroduce a new system for pedestrian detection in rainy or snowy weather.Firstly, we propose a cascaded classification mechanism; and then, in order toimprove detection rate, we adopt local appearance features of head, body andleg as well as global features. Besides that, a specific classifier is designed todetect mirror images in order to reduce false positive rate. The experiments in asingle optical camera based pedestrian detection system show the effeteness ofthe proposed system.

In a pedestrian detection system, a classifier isusually designed to recognize whether a candidate is apedestrian. Support vector machines (SVM) have become aprimary technique to train a classifier for pedestrian detection.However, it is hard to give the best training model which has atremendous effect to the performance of a SVM classifier. In thispaper, we design special code/decode scheme and evaluationfunction for a training model firstly; and then use geneticalgorithm to optimize key parameters which represent the SVMtraining model. Therefore a most suitable SVM classifier can beobtained for pedestrian detection. Experiments have been carriedout in a single camera based pedestrian detection system. Theresults show that the evolutionary SVM classifier has a betterdetection rate; moreover, RBF kernel is more suitable thanpolynomial kernel when chosen in an evolutionary SVMclassifier for pedestrian detection.