Pedestrian Detection Based On DPM Model And Research On KCF Tracking Algorithm

Pedestrians are the main participants in the traffic system.Because of the frequent traffic accidents,how to use the driver assistance system to protect the safety of pedestrians has become one of the hotspot issues.In this paper,the pedestrian detection and tracking are the two aspects of a preliminary study.In the aspect of pedestrian detection,this paper are based on DPM(deformable part model)model and use HOG features of vector quantization for pedestrian detection.In order to adapt to changes in the size of the object,the DPM model adopts the Pyramid of feature and the sliding window method.DPM model use deformable part model mechanism and increase the cost value of the model in the local area of the target component so that deformation adaptability is greatly enhanced.DPM is the best learning algorithm of the traditional machine learning model(not deep learning).However,the DPM model has many disadvantages,such as complex algorithm,poor real-time and poor handling of occlusion problem.In this paper,the real-time performance of the algorithm is discussed.Although Felzenszwalb use cascade detection algorithm to accelerate the improvement of the original,but there is still much room for acceleration effect.In this paper,we use the quantitative feature vector instead of 31 dimensional feature vector,while integrating the cascade detection framework to make improvements.Experiments show that under the same experimental conditions,the detection speed of the improved algorithm is about 8 times that of the cascade detection algorithm,Real time is greatly enhanced.In the aspect of pedestrian tracking,we improve the original CSK algorithm which can not adapt to the change of the target scale.In this paper,we use three scale tools to enlarge or reduce the target template and select the maximum score.then,we update classifier parameters and target template scale.Experiments show that under the same experiment.Improved KCF tracking algorithm have improved in the success and precision aspects.Moreover,when the scale tool P is 3,the FPS is about 47Hz to meet the real-time requirements.