Research On Pedestrian Detection Method In Complex Road Scene

Pedestrian detection is a key technology in the fields of autonomous driving,robot,intelligent video surveillance,etc.It is also an important part of computer vision tasks.In the past decade,from the early traditional manual feature method to the current mainstream deep learning method,pedestrian detection technology has made great progress.However,it is still difficult to achieve accurate,fast and practical detection methods,since the complex changes in the road environment,such as occlusion and light changes.The thesis mainly focuses on the problems of occlusion and all-weather detection,while taking into account the accuracy and real-time.The main research contents are as follows:Aiming at the problem of low accuracy of occlusion pedestrian detection,an occlusion pedestrian detection method based on attention mechanism is proposed.The method first uses the pyramid feature extraction network to obtain the features of different scales and information of the image.Then,based on the feature channel and spatial attention mechanism,a feature-guided attention module is designed to guide the network to focus on the occlusion targets,and this module is embedded into the multi-scale feature fusion process.Finally,unlike the prediction method based on priori bounding box(i.e.anchor)adopted by mainstream methods,this method converts the pedestrian detection problem into a high-level semantic feature detection problem,and obtains the prediction results by heatmaps.Experiments conducted on public datasets verify that this method has advantages in occlusion pedestrian detection and real-time performance.To solve the problem of illumination change in the all-weather pedestrian detection task,especially the problem of high miss rate under low illumination conditions at night,a detection method based on multi-spectral feature fusion is proposed.Since the current all-weather pedestrian detectors lack real-time detection capability,this method first constructs a lightweight feature extraction network to extract the features of visible light images and infrared images,respectively.Then the multi-spectral features are fused through the designed multi-modal feature fusion structure.In addition,the improved low-illumination image enhancement method enhances the visibility of pedestrians in the night visible light image to further imporve night detection capability.Finally,the heatmap prediction method with anchor-free is used to obtain the final results.Through experiments conducted on public datasets,it is verified that the proposed method can better adapt to the all-weather detection environment,especially in the night environment,it has achieved leading detection accuracy and real-time detection capability.