Research On Pedestrian Detection From Traditional Machine Learning To Deep Learning

Pedestrian detection has a wide range of applications in intelligent transportation systems,public safety,robotics and many other fields.However,due to the influence of pose,wearing,occlusion and scale,robust and efficient pedestrian detector is still to be studied,and it is also a research hotspot in computer vision.In the past research,pedestrian detection algorithms are roughly divided into two categories:one is based on traditional machine learning methods,and these methods focus on improving the extraction of manual features,learning of feature classifiers and post-processing;the other type of method named deep learning based on neural networks,this type of method mainly studies the improvement of classification networks.The traditional machine learning pedestrian detection methods are faster,but due to the simple manual features and incomplete post-processing,the actual detection accuracy are not ideal,especially in the case of occlusion or poor illumination conditions.The precision of deep learning objects detection algorithms based on convolutional neural network are better than traditional methods,but due to the characteristics of convolution and pooling operations,the feature map will become smaller and smaller,and the information of some small targets may be covered finally,so the detection for small targets is not robust.Based on the detailed analysis of the theory in machine learning pedestrian detection,this paper has partially improved the problems of the two types of pedestrian detection algorithms.The main research contents are as follows:(1)In order to effectively improve the traditional machine learning pedestrian detection technology in the environments of occlusion or poor illumination,this paper proposed to use the kinect camera to introduce the depth map and produces RGB-D pedestrian dataset,the complementarity between multi-modal data is used to enrich the features,which effectively improves detection accuracy.Based on the ACF(Aggregate Channel Features)people detection algorithm,the experimental results on the RGB-D Person Database show that the accuracy of the proposed method is improved by 16%and the average log miss rate is reduced by 15.9%.(2)The non-maximum suppression algorithm is the core of the post-processing module in pedestrian detection which has a great impact on final detection.In this paper,the classical greedy non-maximum suppression algorithm is partially improved by combining the characteristics of pedestrian,the comprehensive judgment of position and confidence is used to effectively eliminating false detection boxes and keeps the correct detection boxes.The experimental results on the INRIA dataset show that the accuracy of greedy non-maximum suppression algorithm in this paper is improved by 18.85%than the original greedy non-maximum suppression algorithm,and about more than 870 false detection boxes are removed.(3)Due to the low regression accuracy and poor detection performance for the small targets in deep learning one-stage target detection methods,the upsampling inverse feature fusion network is proposed for improvement in this paper.Since the low-level features in the convolutional neural network have richer details and texture information but the receptive field is small,while the high-level features have richer semantic information but more detailed information are lossed.This paper take the SSD(Single Shot MultiBox Detector)as the backbone detection algorithm,the high-level feature layer is upsampled and effectively combined with the underlying features,so that these low-level features have richer semantic information while retaining the original texture information,which is more conducive to the detection of small targets.The comparison experiments results of multiple algorithms on the VOC Person Dataset show that the proposed algorithm has an improvement of 5.97%compared with the SSD detection algorithm,and the accuracy on the small targets set is improved by 6.54%.The mAP value of multi-target detection on the VOC2007 dataset achieved 74.79%.(4)In order to enable the lower layer features to learn more detailed texture information and suppress other noise,this paper designed a segmentation supervised learning network to supervise the learning of low-level features by adding segmentation loss,which can conduct low-level neural networks to learn more feature information.The experimental results on the VOC Person Dataset show that the segmentation supervision mechanism can further improve the detection accuracy by 1.95%,and the accuracy on the small targets set can be further improved by 2.57%.