Research On Real-time Pedestrian Detection Algorithm Based On Depth Features

In recent years,with the maturity of deep learning technology and the innovative application of researchers,the performance of tasks related to computer vision has been greatly improved,especially in the field of pedestrian detection.Pedestrian detection is a sub-task of target detection,Some mainstream object detection algorithms are also suitable for pedestrian detection tasks.And at present,the research of pedestrian detectors based on deep convolution features has become one of the most popular research directions.In this paper,after studying the mainstream one-stage and two-stage pedestrian detectors in detail,improved pedestrian detection models are proposed based on the onestage detector model.The new detector can not only improve the benchmark detection accuracy,but also maintain a real-time detection speed.In order to solves the problems of SSD detectors' poor performance on small-scale targets and inability to handle dense pedestrians,we propose several improvement strategies.First,we adopt a densely connected strategy on the basic feature extraction network of the SSD to fuse shallow features with target features so that the target features can contain more low-level information.Then,during the training process,a fine-grained sampling strategy is used to generate a global prior frame,and a TMMR index to measure the quality of the global prior frame is proposed.Experiments show that after using this strategy,almost all training targets in the entire training set can be matched with corresponding priors,thereby effectively increasing the number of positive samples in the training process and further improving the detection accuracy of pedestrian targets.Finally,we designed an additional loss strategy to make the training data and detection network fully trained and fit,and to a certain extent,solve the problem of a large number of missed detections of downlink human targets in dense scenes.Inspired by the ALF detector,we have studied the reasons why the two-stage detector can have higher detection accuracy,and have added a multi-stage cascade detection module to the improved one-stage detection model.This enables the one-stage detector to improve detection accuracy at a real-time detection speed.The multi-stage cascade detector consists of a basic network and a multistage cascade detection module.First,multi-scale target features are extracted on the optimized basic network.Then,there are multiple detections which are performed on each feature,and the detection scheme on each feature is fed back to the subsequent detection steps.The subsequent detection process fuses the detection scheme with the global prior to generate a larger number of positive sample data,thereby obtaining more accurate detection results during the detection process.Through some comparative experiments on the PRW,Caltech,and VOC pedestrian datasets,it can be proven that the improved pedestrian detection algorithm does achieve the expected detection effect.