Research On Pedestrian Detection Algorithm In Autonomous Driving Scene Based On Deep Learning

Pedestrian detection is an important research branch in the field of computer vision.It plays a very key technical support role in the automatic driving system,which can effectively ensure driving safety and improve the reliability and comfort of automatic driving technology.However,in complex automatic driving scene,pedestrian detection presents many challenges.On the one hand,the pedestrian scales vary greatly in different distances from the on-board camera,it is difficult to detect the small-scale pedestrians with a long distance.Moreover,factors such as illumination changes and background interference also increase the difficulty of pedestrian detection.On the other hand,the occlusion of crowded people and the occlusion of pedestrians by other objects on the road will easily lead to false detection or missing detection.To solve the above problems,this thesis studies and improves the pedestrian detection algorithm based on deep learning.The main work contents of this thesis are as follows:(1)Construct driving scene pedestrian datasets.Firstly,the current mainstream driving scene datasets are introduced in detail.Secondly,the challenging City Persons dataset and SODA10 M dataset are selected according to the comparison results of the diversity of datasets and considering the road traffic differences at home and abroad.Finally,in view of the lack of night image in SODA10 M training set,it is extended by means of homemade night dataset.(2)Aiming at the problem that SSD algorithm fails to detect a large number of small-scale pedestrians in driving scene,a small-scale pedestrian detection algorithm based on multi-scale feature fusion is proposed.First,the input mode of SSD network is optimized according to the characteristics of dataset resolution.Meanwhile,the size and aspect ratio of the SSD default prior box are reset to obtain the prior box that better matches the ground truth box of the dataset.Secondly,in the multi-scale detection stage,shallow high-resolution feature maps are added to improve the detection ability of small-scale pedestrians,and redundant deep feature maps are eliminated to improve detection efficiency.Finally,a feature fusion module is proposed which can make up for the semantic information of pedestrians in shallow feature maps.Experimental results show that the proposed algorithm can effectively reduce the missing detection of small-scale pedestrians,and has good detection effect on different datasets.(3)Aiming at the pedestrian occlusion problem in driving scene,meanwhile,the computing cost and storage space of the model are taken into account,a occlusion pedestrian detection algorithm based on YOLOX-Tiny network architecture and attention mechanism is proposed.Firstly,efficient channel attention module is added into the feature fusion network to enhance the model’s attention to the visible area of occluded pedestrians.Secondly,the boundary frame regression loss function based on distance intersection ratio is used to improve the positioning accuracy of occluded pedestrians.Finally,the structural re-parameterization is used to accelerate model reasoning speed and reduce model complexity.Experimental results show that the proposed algorithm can effectively improve the detection effect of occluded pedestrians,and has good detection performance while ensuring light weight.