Multi-scale Pedestrian Real-time Detection Based On Convolutional Neural Network

In recent years,with the wide application of intelligent video surveillance,driver assistance system and human behavior analysis,pedestrian detection technology has become an important research topic in computer vision.Traditional methods detect fast but lack accuracy,while methods based on convolutional neural network have high precision but cannot meet requirement of detection speed.In this paper,we research the problems of pedestrian detection methods based on convolutional neural network,and based on the previous researches,we propose the improved methods from three aspects: multi-scale,multi-occlusion and real-time pedestrian detection.The main research results of this paper can be divided into the following three aspects:(1)To solve the multi-scale problem caused by different aspect ratios of pedestrians in the driver assistance system,we propose a multi-scale aware pedestrian detection method based on improved full convolution network.Firstly,we introduce deformable convolutional layers in full convolutional network structure to expand the receptive field of feature maps.Secondly,we use cascade-RPN to extract multi-scale pedestrian proposals and introduce discriminant strategy,define a multi-scale discriminant layer to distinguish pedestrian proposals category.Finally,we construct a multi-scale aware network,uses Soft-NMS to fuse the output of classification score and regression offsets by each sensing network to generate final pedestrian detection regions.The experiments have shown that there is low detection error on the datasets Caltech and ETH and our algorithm is better than the accuracy of all detection algorithms in the two sets,works particularly well with far-scale pedestrians.(2)Based on the research of multi-scale pedestrian detection,we find that lower accuracy and speed of our algorithm to detect multi-occluded pedestrian,so we propose a multi-occluded pedestrian real-time detection method based on improved R-FCN.Based on R-FCN,we introduce RoI Align layer to solve misalignments between the feature map and RoI of original images;we improve a separable convolution to reduce dimensions of position-sensitive score maps to improve detection speed.For multi-occluded pedestrians,we propose a multi-scale context algorithm to adaptively output different scale RoIs.In the situation that a lower visibility of the body occlusion,we propose deformable RoI pooling layers to expand the pooled area of the body model.Finally,in order to reduce redundant information in the video sequence,we use Seq-NMS algorithm replaces traditional NMS algorithm.The experiments have shown that there is low detection error on the datasets Caltech and ETH,works particularly well with multi-occluded pedestrians.(3)To solve the problem that parameters of the detection model are too large and cannot achieve real-time detection in the embedded system,we propose a method base on channel pruning to achieve compression and acceleration of model.Firstly,we use channel selection and feature map reconstruction to minimize the reconstruction error of the output feature map.Then,in order to improve the parameter discriminating power of the middle layer channel,we propose a discrimination-loss function to adaptively remove the unimportant parameters in the middle layer and reduce the redundancy of model.Finally,we used a three-stage acceleration method of “training-pruning-fine tuning” to reduce the total test time.By comparing the pruning ratio and the operation time of the algorithm,we reduce the parameter quantity of the training model without guaranteeing the loss of detection accuracy,and apply the algorithm to the embedded system,achieving real-time detection.