A Pedestrian Detection Algorithm Based On Deep Deconvolution Networks In Complex Scenes

Pedestrian detection is widely used in computer-aided driving systems,video surveillance,and robotics development.Due to the influence of many factors,such as uneven illumination,severe obstruction,and extremely low resolution,pedestrian detection in complex scenes still has serious problems such as missed detection and false detection in practical applications.Target detection based on deep learning has been widely applied to big data analysis and processing,and has achieved remarkable results,providing ideas for improving pedestrian detection algorithms in complex scenes.The paper compares the accuracy and speed of traditional pedestrian detection methods and pedestrian detection methods based on deep learning.Combined with the existing problems of pedestrian detection,the paper focuses on the deep learning pedestrian detection and recognition algorithms for small target detection.DSSD with relatively good performance has improved design.Aiming at the resource waste of the residual network(ResNet)in the DSSD detection model,a pedestrian detection network improvement model based on Densenet and DSSD framework is designed.Based on the DSSD network framework,the DenseNet network is used to replace the parameter utilization.ResNet,which is not high,redesigns the replaced DenseNet structure in combination with the principle 'of deep supervision.At the same time,in light of the difficulty of detecting small pedestrian targets in complex scenes,the paper improves the rewinding of small target detection in the network model.Product modules,and redesigned pre-selected boxes with different aspect ratios based on pedestrian characteristics,improved the detection of small-scale pedestrians.Experimental results show that compared with the reference algorithm,the proposed pedestrian detection algorithm significantly improves the detection performance and reduces the resource consumption.In terms of the number of parameters,the number of parameters of the new detection framework proposed by the paper is only 1/6 of the reference algorithm;in terms of speed,under the same hardware conditions,the detection speed of the network proposed by the paper is 45.3%higher than the reference algorithm;In terms of performance,under the same experimental conditions,the network model proposed in the paper achieved 87.84%accuracy for pedestrian detection in complex scenarios,which was 4.07%higher than the reference algorithm.