Research And System Implementation Of Infrared Pedestrain Detection On Edge Devices

As one of the key technologies of intelligent monitoring systems,pedestrian detection can provide real-time information and decision support for people’s lives,transportation,safety,and other aspects.In monitoring systems,low-light environments are important application scenarios,and infrared pedestrian detection has become a research hotspot for nighttime pedestrian detection tasks because it is not affected by changes in lighting.In addition,monitoring systems typically use edge devices for data processing and exchange,but edge devices are difficult to deploy relevant models due to hardware limitations.The key issues in deploying infrared pedestrian detection models include:1)addressing mismatches in public infrared pedestrian datasets and false alarms;2)addressing issues of scale inconsistency and occlusion;3)addressing edge device constraints.To address these issues,this thesis’s research content is as follows:(1)To solve the problem of public dataset scenes,this thesis follows the principles of authenticity,diversity,and challenge,and establishes an infrared pedestrian detection dataset(VOS)for nighttime monitoring scenes.This thesis also statistically analyzes the label distribution,aspect ratio distribution,and other prior knowledge of the VOS dataset.In addition,to address problems such as unclear details,video flicker,and pedestrian ghost reflections caused by the imaging principles of infrared images,this thesis designs an adaptive contrast enhancement method based on image sequences to reconstruct images and verifies the effectiveness of this method through experimental comparisons with other contrast enhancement methods on different detectors.(2)Due to the hardware limitations of edge devices,this thesis uses YOLOv4 tiny as the basic algorithm and conducts research.Specifically,to address the problem that the original K-means clustering is not suitable for non-standard distribution pedestrian datasets,this thesis designs the IKGSDA clustering method to regenerate prior boxes and address the problems of K-means friendly large-scale object boxes,easy to fall into local optimal solutions,and invalid anchor allocation.Although the Mosaic data augmentation method used in the YOLO series can help improve the robustness of the model for pedestrians of different scales and occlusions,it introduces noise samples and leads to a decrease in model performance.Therefore,this thesis proposes a KMPD data augmentation method,which designs four information preservation strategies to improve the fidelity of generated images and avoid ambiguity between generated images and original image annotation criteria.To address the intra-class occlusion problem of dense pedestrians,this thesis integrates other advanced algorithms,designs fusion attention mechanisms,neck networks with down-sampling channels using depth separable convolutions,and introduces repulsion loss to improve the detector’s detection performance and positioning ability for occluded pedestrians.(3)To apply pedestrian detection in practical scenarios,this thesis implements the deployment of pedestrian detection models and the development of nighttime monitoring systems.This thesis chooses the MaixSense development board as the edge device and performs model conversion,operator compatibility,and other operations to ensure that the model can run in real-time on the edge device.The pedestrian monitoring system can perform all tasks well and execute all requests in pressure test with a delay control of less than 800 milliseconds,without accumulating delay problems,and can meet actual application requirements.