Research And Implementation Of Pedestrian Detection Algorithm Based On Convolutional Neural Network In Crowded Scenes

Pedestrian detection is a very important task in computer vision,a key component of various real-world applications,and plays an important role in autonomous driving technology,security,and robotics.Pedestrian detection is a specific object detection task for detecting pedestrians in images and accurately predicts the bounding box containing each pedestrian.In recent years,with the rise of deep convolutional neural networks,the performance of pedestrian detectors has been rapidly improved.However,due to the occlusion between pedestrians in crowded scenes,the detector may incorrectly consider the occluded pedestrians as a whole at this time.Therefore,the accuracy and speed of current pedestrian detection algorithms still cannot satisfy the needs of practical applications in crowded scenes.To address both accuracy and speed problems,this paper proposes a high-quality proposal features generation for crowded pedestrian detection algorithm and an adjacent feature complementary for crowded pedestrian detection algorithm,respectively.The main research works and innovations are as follows:(1)High-quality proposal features generation for crowded pedestrian detection algorithm.In order to improve the accuracy of pedestrian detection in crowded scenes,this paper proposes a dual-region feature generation algorithm.Firstly,visible regions with less occlusion are introduced and low-precision proposals are generated for both the full-body and visible regions respectively.Then,proposals are respectively selected from the two kinds of proposals mentioned above to match in pairs,so as to guarantee a strong correspondence in information between the two proposals.Afterwards,the successfully matched proposal features are fused by Selective Kernel Feature Fusion to generate high quality proposal features.Secondly,Paired Multiple Instance Prediction is performed on the fused features to generate multiple prediction branches,and each prediction branch generates full-body and visible prediction box.Finally,Paired Non-Maximum Suppression is applied to the prediction boxes to reduce the false positives.Experiments on the CityPersons and CrowdHuman datasets show that this algorithm can further improve the pedestrian detection accuracy in crowded scenes.(2)Adjacent features complementary for crowded pedestrian detection.In order to improve the balance between accuracy and speed,an adjacent features complementary algorithm for crowded pedestrian detection is proposed on a one-stage detector.This algorithm firstly invokes a deep expansion convolution module to obtain deeper features.Next,a hierarchical feature extraction module is designed to extract diversity feature information.That is,the multi-scale feature extraction and channel attention mechanism are mainly used on the high level features to extract the contextual information between features,and the spatial attention mechanism is applied on the low level features to filter the background information.Finally,the adjacent pyramid feature Integration module is proposed to improve the detection results by aggregating the associated features of adjacent layers.Experiments show that this algorithm can increase the detection speed of the algorithm while maintaining the accuracy improvement.This thesis contains 18 figures,14 tables and 84 references.