| As pedestrian is an most active,complex and important element in application scenarios,pedestrian detection and tracking technology has become a popular research direction in China and abroad,which is widely used in emergency rescue and disaster relief,epidemic prevention and control,public safety monitoring,intelligent transportation and assisted driving and other fields.Due to the influence of apparent changes of pedestrian in appearance,the interference of surrounding environment and other special factors,traditional machine learning algorithms have been unable to meet their performance and real-time requirements.This paper designed a fast and highly accurate pedestrian detection and tracking algorithm based on the method of multi-feature fusion.The main research contents are as follows:Under the complex background conditions,the single feature was insufficient to describe pedestrian resulting in low accuracy,a fast pedestrian detection algorithm with multi-feature fusion is proposed.First,three pedestrian feature vectors were in blocks and connected in series which were to the feature of Histograms of Oriented Gradients,the improved feature of Color Names and the feature of Local Binary Pattern.Then the Principal Component Analysis was used to reduce dimensionality for the problem of high dimensionality of the fused feature vectors affecting pedestrian detection speed.On the INRIA public pedestrian data set,an adaptive increase algorithm based on support vector machine for weak classifiers was proposed in order to do training and testing.In addition,a pedestrian detection experiment platform was designed according to the actual application requirements,which could quickly implement the purpose of the statistics on the number of pedestrians and the detection time.The results of the experiment showed that multi-feature with complementary relationships can effectively improve the pedestrian detection rate,and the dimension of the improved fusion feature vector was reduced by 48% compared with the original.The test of the experimental platform reached an accuracy rate of 93.2%,with an average detection time of 0.28 s.It shows that while the accuracy of the algorithm had improved,the detection speed had also improved correspondingly,the algorithm in this paper was better than the other four algorithms.Based on the pedestrian detection algorithm,the feature fusion idea was applied to the pedestrian tracking algorithm,and a feature fusion and multi-scale of fast tracking algorithm kernel correlation filter of pedestrian is proposed.A new confidence measurement method for response graph was proposed.And this method combined two indicators of the Peak Side-lobe Ratio and the Average Peak-to Correlation Energy into a confidence evaluation index by setting the joint confidence factor.The new feature was weighted by three features which automatically adjusted the fusion weight coefficient.So it can fully utilize the advantages of different features in different scenarios which solved the problem of fusion of fixed weight coefficients in the kernel correlation filter algorithm The performance of the filter algorithm was improved.Then the scale pool method was used to solve the problem that the fixed detection window cannot adapt to the change of the target size.Finally,the model was updated by setting the joint confidence threshold as the updating condition,in order to reduce the cumulative error brought by frame-by-frame updates and the number of model updates.This paper used the pedestrian sequence on the OTB-100 public dataset to conduct experiments.Comparing the algorithm in this paper to the four methods,the results proved that the algorithm in this paper can better reduce the influence of factors such as illumination variation and object deformation.The average speed of the experiment reached 26.2 frame /s.The average accuracy of the pedestrian tracking sequence was 83.4%,and the average success rate was 77.8%,which were better than the comparison algorithm. |
PDF Full Text Request