Research On Pedestrian Detection Based On Feature Fusion And Deep Learning

Pedestrian detection is a popular research topic in the field of artificial intelligence,which has broad significant applications such as intelligent video surveillance,driver assistance system and robots of intelligence.However,due to the influence of complex background and the variability of multi-pose or body,current pedestrian detection methods can hardly meet the practical requirements.Therefore,it is necessary to conduct further research and exploration.Feature extraction and classifier training are two main parts for pedestrian detection.How to design a robust feature extraction method that is not influenced by the environment and how to choose a classifier with strong discrimination ability become a crucial factor to determine the performance of pedestrian detection.Feature extraction methods mainly can be divided into hand-crafted and learning-based feature extraction approaches.In this paper,we mainly focus on the feature extraction.We develop a new feature representation method and a deep network model to extract novel hand-crafted features and deep learning features respectively.Besides,a coarse-to-fine pedestrian detection framework and a new data augmentation technology are introduced.The main contributions of our work are as follows:The LDCF detection method generates more false positive windows,so this paper proposes a novel detection approach,which is based on convolutional channel features and color self-similarity features.A hierarchical coarse-to-fine detection strategy is used,in which we apply LDCF method to achieve candidate windows firstly.Then convolutional channel features and novel self-similarity features are extracted to represent these windows and are trained by Adaboost algorithm for fine-grained classification to eliminate false detection windows.Experimental results show that such method is able to reduce the number of false positive windows effectively.The detection rates can increase 2.81% and 3.85% in the INRIA and Caltech datasets respectively.In order to improve the discrimination capacity of the LDCF classification model and pick up speed of detection process,a faster multi-scale pedestrian detection method is proposed,which is based on the novel data augmentation skill.We first scale each training sample to augment the dataset.It can improve the learning capacity of the classifier and reduce sensitivity to small noises injected in the ground truth window.Then a faster multi-scale feature pyramid model is applied in the detection process to compute feature channels with different scales quickly,which can reduce the complexity of computing.Experimental results show that the proposed method can reduce the miss rate effectively.It should be noted that we also combine the coarse-to-fine strategy with this method.It achieves the state-of-art detection results among approaches without using the deep learning framework on the INRIA and Caltech datasets.In order to improve the effectiveness of the coarse-to-fine detection procedure,we propose a deep convolutional neural network based detection method,in which the deep network model and regression model are used to optimize and polish the coarse-to-fine strategy.Firstly,we fine-tune the VGG-16 network to train a pedestrian classification modeland use the correlation between position and width of pedestrian bounding box to learn a prior regression model.Then we use the regression model to eliminate some false detection candidate windows stem from LDCF detection model.Besides,high-level deep features are further extracted from remaining candidate windows for fine-grained classification.We can achieve more accurate detection results.Experimental results show that the detection performance can be further improved.In particular,compared to LDCF method,the log-average miss rate of our proposed approach reduces from 24.80% to 11.82%,achieving performance gains by 12.98%.Although our method is still inferior to other best deep learning based approaches,our detection speed is superior to them.Therefore,our method can achieve better tradeoff between accuracy and detection speed.