Research On Efficient Object Detection Methods In Family Clutter Scenes

Problems such as complex and changing home environments and variable object morphology pose a great challenge to target detection.The existing algorithms cannot achieve stable fusion of object features,and suffer from target misdetection and omission problems,which affect the downstream sensing tasks of home service robots.In addition,the algorithm has the problem of being computationally intensive,which makes it difficult to meet the requirements of industrial applications.Thesis conducts research on the above problems and propose a global and local feature fusion complex object detection method based on the self-attentive mechanism,which realizes the global and local feature fusion of observation-constrained targets and achieves stable detection of complex target objects in home clutter environment,and proposes an attention mechanism acceleration method based on the locally sensitive hash algorithm,which realizes the reduction of algorithm computation and memory overhead.The research for thesis is as follows:(1)A self-attentive local and global feature fusion method based on feature map scale invariance is proposed for the problem of limited feature extraction of target objects in ground close family clutter environment.Firstly,a object detection model framework based on inverse operation to keep the scale of feature map invariant is constructed,and the global sparse features of the target are obtained by using convolutional space mapping to ensure the invariant feature map perceptual field.Secondly,for the image information fragmentation problem,a feature mapping method based on window interference is proposed,and the depth-separable convolution is used to achieve simultaneous encoding of location information,which ensures the interaction and completeness of features and the accuracy of location information of image feature sequences.Finally,to address the feature loss problem,multi-scale features are constructed using an attention-based contribution scoring method,and the residual structure is used to learn differentiated features to achieve the improvement of the network’s ability to pay attention to target objects and the environment.(2)A fast attention mechanism with a priori learning capability is proposed for the attention model arithmetic and memory overhead problems.Through the feature screening method based on sparse inflated attention,sparse inflated masks based on pattern combination and pooling combination are used to screen the features with high relevance,reduce the feature sequence while alleviating the information fragmentation problem,and then through the feature classification method based on locally sensitive hash,the hash value is used as a priori to classify the token types and achieve the tight correlation of local attention through the intra-class calculation of attention.Through the inter-class association of attention,the global information is sparsely linked to ensure the information flow and reduce the attention redundancy,and then the model achieves the fast extraction of features within the tight correlation range to achieve the purpose of reducing the model computation and memory overhead.(3)In thesis,a object detection system is designed to verify the effectiveness of the algorithm in line with real home clutter scenario applications.In thesis,we first construct the household clutter scene dataset and design different occlusion levels to verify the algorithm performance.Using mainstream object detection evaluation indexes to compare with the benchmark model,it is verified that the algorithm in thesis can effectively extract monotonic features,achieve fast convergence of the model and fast extraction of key information in a tight range,and reduce the impact of interference factors on detection accuracy.By comparing the model computation and number of parameters,it is demonstrated that the algorithm in thesis contributes to the optimization of computation and number of parameters without degrading the object detection effect.