| At present,the pedestrian detection technology based on deep learning has been greatly developed and is widely used in video processing fields such as intelligent monitoring,automatic driving,pedestrian analysis and so on.However,due to the large amount of parameters,high algorithm complexity and heavy dependence on hardware computing power of the deep network model,the existing pedestrian detection network is difficult to deploy in practice.Therefore,with the support of National Natural Science Foundation of China,this thesis conducts research on pedestrian detection technology of lightweight network.The innovative research results obtained and the major work completed are as follows:(1)Inverted residual lightweight pedestrian detection algorithm based on LGC.Existing mainstream detection models usually increase the depth and width of the network to acquire better detection results,but it is unamiable to suffer from parameters increasing,detection rate decreasing and so on.Therefore,an efficient lightweight convolution model was proposed by referring to the lightweight idea of Ghost convolution and group convolution,which is to balance the detection accuracy and speed,so as to obtain more feature maps with fewer parameters.CSPDarknet53 that was redesigned with the above convolution and an inverted residual structure was introduced to generate an inverted residual feature extraction network to improve the global feature information extraction capability of the model.On this basis,the inverted residual feature extraction network was used as the backbone network of YOLOv4,and depthwise separable convolution was used to reduce the parameters.To improve the overall performance of the algorithm,an inverted residual target detection algorithm was proposed.Experimental results show that the proposed algorithm has prominent advantages in the number of model parameters and detection speed under the premise of similar detection accuracy.(2)Lightweight pedestrian detection based on multi-scale information and crossdimensional feature guidance.Aiming at the problem that the detection accuracy of pedestrian detection in complex road environments,a lightweight pedestrian detection algorithm based on multi-scale information and cross-dimensional feature guidance is proposed.This paper takes the high-performance detector YOLOX as the basic framework,a multiscale lightweight convolution is constructed and embedded in the backbone network to obtain multi-scale feature information.An end-to-end lightweight feature guide attention module is designed,which fuses spatial information and dependent correlation information by cross-dimensional channel weighting,and then guides the model to focus on visible region of pedestrian targets.Construct a feature fusion network using a depthwise separable convolution with a large receptive field,so that the model can reduce the loss of feature information while reducing the amount of parameters.The experimental results confirmed that the proposed algorithm can greatly enhance the detection accuracy in scenes such as complex environment and dense occlusion,and has better robustness and real-time performance.(3)Lightweight pedestrian detection algorithm for multiple scenes.Aiming at the problems of large volume and high amount of computation in the practical application of the existing pedestrian detection models,which are difficult to deploy on the mobile devices with limited resources,a lightweight pedestrian detection algorithm for multiple scenes is proposed.Firstly,a deep and shallow feature fusion network is constructed to learn the texture feature of multi-scale pedestrians;Secondly,Average pooling is added in the cross-dimensional feature guide attention module to reduce the loss of feature information in the pooling process.Finally,the redundant channels in the model are trimmed by the pruning strategy to reduce the algorithm complexity.In addition,an adaptive Gamma correction algorithm is designed to reduce the influence of external disturbances such as illumination and shadows on the detection results in multiple scenes.The experimental results show that the proposed method can compress the model volume to 10 MB and the processing speed can reach93 frame/s under the premise of similar detection accuracy,which outperform the current mainstream methods. |
