Research On Pedestrian Detection Based On Saliency And Parts Modeling With Model Recommendation

Pedestrian detection technology applied in the fields of computer vision,such as video surveillance,behavior analysis,automobile driver-assistance,robot control,etc.,enjoys extensive application prospects and research values.In recent years,there has been lots of research achievements on pedestrian detection,but many problems remain unsolved.Further study is required because of human intraspecific difference and complexity of the scene.This paper studies the pedestrian detection technology in three aspects,such as preprocessing,classifier based on parts modeling and model recommendation for adaptive scenarios.The major contents and contributions are listed in the below:(1)Salient region detection is convenient for many computer vision applications and becomes the preprocessing of pedestrian detection.There is a lack of effective integration between the heteroid features and intrinsic and extrinsic cues with dependence on prior assumptions.Aiming at this problem,this research proposed a framework for salient region detection with multi-feature fusion without high-level semantic hypothesis.Firstly,the representations of statistical texture for superpixels as the extrinsic cues were constructed to generate salient prior map.Then,multi-channel color features were constructed to generate salient map by least square method.At last,the color contrast feature was introduced to form the final salient map with information entropy optimization for noise.The experiment was carried out on three popular benchmark datasets such as MSRA,ECCSD and PASCAL-S.Results on analyzing the accuracy and complexity demonstrate that our method is competitive in comparison with the methods based on prior assumptions,such as DSR,MC and RBD.Comparing to DSR and DRFI,there is a obvious reduction in computing complexity.The performance of pedestrian detection is improved by using the technology of salient region detection.(2)The part-based detection methods in the pedestrian detection have excellent performance and have a strong adaptability in posture change of human body.But these methods exist sensitive problems in model initialization and occlusion.In order to solve the problem of initialization sensitivity for non-convex optimization of model training with parts latent variables,this research proposes a model initialization algorithm based on face priori information.By applying face detection technology to model initialization,the accuracy of model training was improved with high-level priori information instead of image intrinsic characteristics.In order to solve the problem of parts occlusion,a pedestrian detection algorithm was proposed based on occlusion modeling.Considering the disadvantage of LSVM method for mining hidden information,the research proposed a two layers classifier based on the deformable parts model and established conditional random fields model for occlusion status.For learning model parameters,the research used the stochastic gradient descent and belief propagation algorithm to optimize objective function of the conditional random fields.The experimental results on PASCAL VOC data set show that the proposed approach achieves good performance for occlusion problem,and effectively improves pedestrian detection performance for deformable parts model.(3)A UDN model combining the convolution features from deep learning with parts modeling is performed well in pedestrian detection.However,the UDN model is still a strict part model,and the latent training of the part greatly affects the judging ability of pedestrians.This research proposed a pedestrian detection method which trained viewpoint selection model by deep reinforcement learning and simulated human vision to search local key parts based on visual attention mechanism.The method generated focused images by viewpoint selection model to search and overlay them together for key areas,and detected pedestrians by detecting network.Then,information entropy was computed for measuring the reliability of the result and optimized the viewpoint selection model as a reward for deep reinforcement learning.The collaborative iterative training with viewpoint selection model and detection network,made the method having a strong ability for searching and detecting local key areas,and reduced the influence for posture changing with body deformation and occlusion.Comparing with classic part-based pedestrian detection methods on public pedestrian detection data sets,such as Caltech and ETH,the experimental results show that the proposed method can effectively improve the pedestrian detection accuracy.(4)Different pedestrian detection models have different performance in different scenes,and the precision and complexity are two contradictory factors.Balancing precision and complexity and giving full play to respective advantages of models are helpful for the application of the pedestrian detection.Considering the feature of the scene,the research introduced recommender system into pedestrian detection and proposed a pedestrian detection method depending on complexity with model self-adaption.This method modeled scene features as a probe response and made weighted scores for precision and complexity of models as a score matrix.Then,the method mined the inner relation between scenes and detection models by collaborative filtering algorithm.According to the weights for complexity and characteristics of the input image scene,the optimal model for pedestrian detection was dynamically selected in the detection model set.Experiments show that the efficiency is improved significantly and play full advantage of low complexity models in some specific scenes.