Research On The Method Of Vision Pedestrian Detection Based On Transfer Learning

With the rapid development of artificial intelligence,pedestrian detection has been widely used in many fields of computer vision,such as video surveillance,robot navigation,vehicle driving assistance system and so on.In the case of different distribution of multi-scene data,when a detector with excellent training performance is applied to a new environment,it is vulnerable to the interference of shooting angle,complex environment and pedestrian posture,resulting in a significant decline in its performance.Inspired by the successful application of transfer learning,deep neural network,sparse coding and background subtraction method in pedestrian detection,this paper proposes a visual pedestrian detection method based on transfer learning.the research contents of this paper are as follows:1.A pedestrian detection algorithm based on feature optimization and background subtraction is proposed.First,a general detector is designed to optimize the traditional Histogram of Oriented Gradient(HOG)with Support Vector Machine(SVM)algorithm,combine HOG and Local Binary Pattern(LBP)for feature extraction,and detect the initial sample set of the target scene.Then,an improved Gaussian Mixture Model(GMM)background subtraction method is used to detect the moving regions in the target scene.The experimental results show that the improved background subtraction method can quickly and effectively detect long-distance pedestrian targets,thus enriching the target samples.2.A transfer learning method based on sparse coding is proposed.In the process of sample selection,sparse coding chooses irrelevant or very different atoms in order to accurately reconstruct the original signal,which makes the classifier unable to effectively learn the classification surface.By adding local constraints,the method makes the coding of similar features similar,and realizes the transfer of a small number of source samples suitable for target scene.The weights of different samples are reallocated by improved structured sparse coding,and the final classifier is retrained.Experiments on multiple test sets show that the algorithm can adapt to different scenarios and has strong robustness.3.A Deep Sparse Auto-Encoder Network(DSAEN)is proposed to solve the problem that the traditional convolutional neural network takes a long time and has large redundancy.First,by studying the inherent properties of pedestrians,the Norneighboring and Neighboring Features(NNF)feature descriptor are optimized to capture the symmetry features of pedestrian contours and the differences inner pedestrians,pedestrians and backgrounds,which provides abundant supplementary information for HOG and LBP.Then,from the aspects of activation function design,loss function design,and sparse design,a new objective function is constructed to learn the deep auto encoder network,and obtain the structural features of the image.4.A pedestrian detection method based on deep network transfer learning is proposed to solve the over fitting problem of a small number of training sample learning networks.First,the network model S-DSAEN in the source domain is designed.Then the network trained in the source domain is transfered to the new network in the target domain,and the network structure is optimized and fine tuned,and the spatial information is fully utilized to train the classifier.Experiments show that the post transfer deep sparse auto encoder network saves the training time and computing resources of the deep network model,and improves the accuracy and detection speed significantly.