Rotation-invariant Human Detection In The Rescue Scene

Rescue robots are playing an increasingly important role in post-disaster rescue efforts.Visual sensors that have been carried on robots widely most closely resemble the way in which humans perceive the environment and can provide abundant environmental information.How to realize visual detection for human bodies in the recue scene is one of the key technologies to enhance the autonomy of rescue robots and promote the efficiency of searching and rescuing work,which has an enormous theoretical and practical meaning.Different from pedestrians appearing in upright poses such as standing or walking,the human bodies in the rescue scene have more varied postures,the most fundamental specificity of which is the uncertain planar rotations between the captured object and the imaging equipment.This dissertation studies the rotation-invariant human detection from the aspects of visual feature,classification strategy and rotation-invariant semantic description.The main research achievements are as follows.1.On the basis of Histograms of Oriented Gradients(HOG),a new visual feature is proposed to facilitate the rotated object detection.First,for the gradient orientation calculation,we take the dynamic local coordination system,which varies with the pixel position,instead of the fixed global coordination system to describe the pixel's gradient.This modification can make the pixel's gradient orientation invariant in the image rotations.Second,we map the detection window in the polar coordinate system with its center as the origin and adopt the sector-ring as the spatial shape of sampling cells.This ensures that the image rotation can be simplified into the cyclic shift of feature vector.Third,the computations of gradient interpolation and Gaussian block-weighted window in HOG are adjusted to adapt to the new feature.Due to the sector ring shape of sampling units,the new feature is called Sector-ring HOG(SRHOG).To prove the new feature's effectiveness,SRHOG is tested with linear Support Vector Machine(linear SVM)on INRIA pedestrian dataset,INRIA128×128 pedestrian dataset and VD01 victim dataset.Experimental results show that SRHOG has the comparable pedestrian discrimination with HOG and performs better than other advanced pedestrian detection methods in victim detection,which demonstrate SRHOG's superiority in rotated object detection.2.Combining Boosted Random Ferns(BRFs)with SRHOG,the rotation-invariant detection strategies for complicated objects are studied.First,we define the Local Binary Features(LBFs)and Ferns in SRHOG feature domain,and choose the most discriminative ferns by the Real Adaboost algorithm to obtain the final classifier.Based on the method,i.e.SRHOG-based BRFs,two different rotation-invariant detection strategies are proposed.The first strategy depends completely on the characteristics of SRHOG and detects the test image repeatedly with the feature's cyclic shifts replacing the image rotations.The second strategy introduces an orientation-estimation step and detects the test image from the predicted orientation.Besides,we perform coarse detection at beginnings of two strategies to filter out some background subimages.In order to facilitate the comparison with other researches,experiments are conducted in Freestyle Motocross public dataset.Experiments contain the orientation-specific motorbike detection,which is for validating the discrimination of SRHOG-based BRFs,and the rotated motorbike detection,which is for demonstrating two detection strategies' abilities to address planar rotations.Results show that our methods can obtain similar performances with other state-of-the-art methods at a low computation cost and are more suitable to the environment with limited computing capability.Besides,in our two detection strategies,the first performs better in finding rotated object and the other can provide more accurate orientation predictions.3.Combining Rotation-Invariant HOG(RIHOG)and Bag of Visual Words(Bo VWs)model,the rotation-invariant semantic description method for complicated object is studied.We learn from HOG and split the global description into low-level features calculation and middle-level semantics extraction,which are realized by RIHOG feature and Bo VWs model respectively.The semantic description method is called RIHOG-Bo VWs and can enhance the discrimination by catching the object's local characteristics.In the experiment stage,RIHOG-Bo VWs is tested with linear SVM on Freestyle Motocross dataset.Results show that our method can achieve comparable effect with HOG in the orientation-specific motorbike detection and perform better than other features in the rotated motorbike detection.This demonstrates the RIHOG-Bo VWs' sufficient discrimination in object detection and its ability in addressing planar rotation.Besides,RIHOG-Bo VWs has the good translation invariance and can improve detection efficiency significantly.4.To evaluate the above detection methods in rotation-invariant human detection,we establish two recorded datasets of human bodies in rescue scenes by imitating the perspectives of unmanned aerial vehicle(UAV)and ground robot.These two datasets are called VD-Drone dataset and VD-Robot dataset respectively.Due to the morphology difference of human bodies between two datasets,methods have different performances in VD-Drone dataset and VD-Robot dataset.For VD-Drone dataset,linear SVM+RIHOGBo VWs and linear SVM+SRHOG have the similar detection effects.For VD-Robot dataset,severe perspective distortions influence the detection and linear SVM+SRHOG achieves the highest Equal Error Rate(EER).Besides,among all detection methods,linear SVM+RIHOG-Bo VWs has the fastest detection speed and its consuming time is about half of HOG.In conclusion,aiming at the rotation-invariant detection for human bodies in recue scenes,this dissertation proposes one visual feature SRHOG,two detection strategies using SRHOG-based BRFs,and one rotation-invariant semantic description method RIHOGBo VWs.We also establish two recorded victim datasets under the perspectives of UAV and ground robot to demonstrate their outstanding effects.This research can promote the application of computer vision in rescue efforts and improve the emergency rescue capability of our country.