Research On Adaptive Pedestrian Detection Algorithm In Complex Scene

Convolutional Neural Network(CNN)completely replaces the manual by the machine to obtain the image features of the input layer through the convolution operation,which improves the recognition ability of the input data.Then,the algorithm designed in this thesis uses a very mature convolutional neural network framework to design pedestrian detection algorithms.For the background interference between pedestrians and the occlusion between pedestrians,the pedestrian's head is used as the detection object,which reduces the feature extraction of the detected object.In order to obtain a better detection effect and speed up the calculation speed and accuracy in the pedestrian detection process,the region proposal network(RPN)in Fast R-CNN with robustness can be used as a better pedestrian detector,but its classifier May degrade performance during detection.Therefore,this thesis improves the target detection network based on the robust Fast R-CNN,and fuse the image feature map obtained by the convolution layer to obtain more abundant and precise samples.Experiments show that the improved method in this thesis can accurately and effectively improve the efficiency of pedestrian detection.The main work of this thesis:(1)Image fusion has different imaging principles,so the obtained image information is complementary.Then the combination of Fast R-CNN and image fusion technology enhances people's cognition and understanding of the original image.In order to obtain better gesture perception,the visual saliency extraction method can highlight the significant information of the source image.(2)This thesis proposes a new weight mapping construction process based on visual saliency features,which can integrate the important visual information of the source image into the fusion image.The average layer is used to obtain the base layer and the detail layer.The fusion base and detail layers use the average fusion and KL transform fusion rules respectively.Finally,the final image is obtained by fusing the base layer and the detail layer.(3)Experiments show that the method of image fusion can obtain high-quality area and texture features in the source image,and its recognition results exceed the current mainstream detection methods in subjective and objective analysis.Aiming at the serious occlusion between targets in dense crowds,the head area of pedestrians can be analyzed at the detail layer and the base layer through multi-scale decomposition during the fusion process to predict the distribution and the possibility of pedestrians in the image.(4)The experimental verification of the algorithm has designed different algorithm structures for comparing the experimental results,and again used different parameter network structures for comparative experiments to verify the improved pedestrian detection algorithm from multiple angles,so as to obtain a comparison with traditional pedestrian detection.The method is more optimized in the analysis of recognition efficiency,accuracy and image details.