Vehicle Detection Algorithm Research Based On Complicated Background On Earth

Vehicle recognition based on complicated on earth background is widely used in military fields,such as battle zone reconnaissance and precision strike,and civil fields like intelligent transportation and environmental monitoring.This paper studies on the vehicle recognition method in images with complex background which collected by airborne optoelectronic pod.According to the shortcomings of existing algorithms,a method of vehicle recognition based on machine learning is proposed.Firstly,the image ROI is extracted by the saliency detection.Then,the improved HOG features of ROI are extracted.Finally,these features are recognized by the random forest classifier.Therefore,the ROIs which contain vehicle are selected and the work of vehicle recognition is finished.The main work and achievements of this paper is as follows:?1?According to the large amount of redundant information like background clutter and noise in images from optoelectronic reconnaissance,a selective background prior for saliency detection is proposed.The existing saliency detection algorithm cannot deal with the image which has no salient object or the object located at the boundary.And the salient object may be incomplete.In this paper,by calculating the background probability based on boundary connectivity,the background templates are selected.Then,the edge and color saliency map is calculated by geodesic distance from background in edge and color features,respectively.And the edge features is calculated by edge detection.By combining the saliency cues with an optimization problem,the ROI of image is extracted.Comprehensive experiments demonstrate the superiority of the proposed algorithm over the state-of-the-art methods in precision,recall and so on.?2?According to the ignoring of color information in HOG and the high dimension of HOG,a low-dimensional SSHOG feature based on the image segmentation and SLPP algorithm is proposed.By using the edge and color information of the image segmentation,the projection weights of the HOG is modified to increase the distance between the target's and background's features.Then,the dimension of SHOG is reduced by using the SLPP algorithm,and the redundant features are removed.Experiments on complicated background on earth images show that the description of SSHOG is improved by combining the edge and color information of region and reducing the dimensions.The accuracy of SSHOG in vehicle recognition from KCRF classifier is 16.03%higher than HOG feature.And theF₁-measureincreases 8.79%.The classification effect of SSHOG is better than existing feature.?3?According to the shortcomings of random forest easy to over-fitting if the sample has large noise and difficult to calculate if feature is continuous value,the KCRF classifier based on local linear KNN sparse representation and random forest is proposed.In order to train the decision tree,K means is used to divide the sample data.And in the prediction,the sample feature is used as a dictionary,and the spares representation of the predicted feature is used to classify.Therefore,the recognition of vehicle targets from ROI in saliency detection quickly and accurately.The results of experiments show that the accuracy of KCRF classifier is 2%higher than the RF classifier.