Study On PCANet For Pedestrian Detection

Pedestrian detection has a wide application value in the fields of vehicle auxiliary driving,human behavior analysis and intelligent robot.In recent years,pedestrian detection has also been gradually applied to aerial images,victim rescue and other emerging fields,which makes pedestrian detection widely concerned by academia and industry.The characteristics of pedestrians both rigid and flexible objects,i.e.pedestrian detection in natural environment are vulnerable to complex background,occlusion and low resolution of non-ideal factors,and the pedestrian itself also has the unstable factors such as the diversity of posture,the body and the diversity of clothes.Point to these influence factors,in this paper,the deep learning theory and the detection of the significance of the object are introduced,and the methods of strong robustness,detection effect and fast detection speed are studied.The concrete research content is as follows:First of all,a novel effective deep learning framework called Multi-Scale Principal Component Analysis Network of Spatial Pyramid Pooling(MS-PCANet-SPP)is proposed in this paper to overcome the challenge of detecting pedestrian accurately and robustly.In the representative Principal Component Analysis Network(PCANet),principal component analysis is used in two convolutional layers,followed by binary hashing in the non-linear layer and block-wise histograms in the feature pooling layer.Then in order to obtain the saliency feature of the image,in the feature pooling layer,the multiscale block-wise histogram features are extracted by the spatial pyramid pooling method.In the output layer,it is well known that high-level features which are learned from latter stage tend to be more global and low-level features which are learned from early stage tend to be more local.Therefore,the output of each stage is stacked together as one final feature output of this paper's model so that the feature representations contain both holistic abstract information and local specific information.The purpose of combining feature representations from multiple stages is to provide multi-scaled features to the classifer.MS-PCANet-SPP is a deep feature extraction network.It needs to extract a large number of candidate regions when detecting multiple pedestrians and large pedestrian images.However,many candidate areas do not contain pedestrians.Extracting feature from these candidate regions will lead to a waste of detection time,which will affect the real-time performance of detection.Therefore,based on the above research,this paper further studies saliency detection algorithm,and proposes a pedestrian detection algorithm combining Binarized Normed Gradients(BING)and MS-PCANet-SPP.First,a series of cascaded support vector machine classifier is trained to predictthe candidate area of pedestrians according to the BING features of the pedestrian in the image and then use MS-PCANet-SPP to extract the feature of candidate regions to determine whether the candidate area is pedestrian or non-pedestrian,so as to achieve accurate pedestrian detection.The experimental results show that,the proposed MS-PCANet-SPP can achieve better pedestrian detection.It not only achieves better detection results,but also has strong robustness to pedestrian detection with non-ideal factors.In addition,the pedestrian detection combined BING and MS-PCANet-SPP can not only maintain high detection rate,but also have faster detection speed.It effectively alleviates the problem that the detection precision and the detection speed are difficult to be balanced in pedestrian detection.