Research On 3D Small Object Detection Method Based On Attentional Feature Enhancement

Three-dimensional object detection,a key technology in autonomous driving task,is responsible for identifying the absence of specific objects from the environmental data collected by on-board sensors and returning its location in the scene.However,there are some objects with small scales in the scenes,and it is difficult for the sensor to acquire sufficient point clouds at these small objects.Limited by the scale and number of point clouds,it is restricted for the network to extract sufficient features from the small objects,which damages the detection performance and thus affects the application of 3D object detection in various technologies.The current research difficulties of small object detection mainly lie in that the small objects have too few points to provide sufficient feature representations;the distribution of small and large scale objects is uneven,and the acquire of rich feature representation by 3D object detection methods easily leading to complex model computation.In summary,a 3D small object detection method based on attentional feature enhancement was proposed,which mainly focus on the detection problem of small objects in3 D scenes.The main work is outlined as follows:(1)An attention enhancement-based multi-scale small object detection method was proposed.Firstly,since the existing 3D backbone networks usually applies multiple upsampling operations to recover the feature resolution for obtaining multi-scale features,an feature pyramid network was proposed that can reduce the computation of up-sampling and achieve efficient hierarchical feature extraction based on few point clouds.Then,based on the attention mechanism,the contextual association between objects in each hierarchical feature was explored,enhancing the feature representation of small-scale objects and suppressing interference from other scales.Thus,the attention of small objects was enhanced and the network can predict objects at different scales in a hierarchical manner.Extensive experiments on the KITTI dataset show that this method improves the overall mean average accuracy by 5.17% and 2.18% over the benchmark method for the small objects of pedestrian and cyclist,respectively,and the detection speed of this method is also 6.48 Hz faster than the benchmark method.(2)An attention-guided feature fusion method for small object detection is proposed.To obtain rich feature representation,the contextual features learned by the backbone network were encoded into the grid representation,while the point-wise spatial information of the native points and the geometric information of the proposal was learned.Then,based on the attention mechanism,the grid-wise contextual information and point-wise spatial knowledge was adaptively integrated,and the global association among grids is further modeled to enhance the representation of the key information of small objects.Extensive experiments on the KITTI dataset show that this approach improves the mean detection accuracy of the benchmark network by 2.07% and 0.91% on the small objects of pedestrian and cyclist,respectively.And the experiments on the Waymo dataset show that this approach improves the average mean detection accuracy of the benchmark network by 3.58% on small pedestrian objects.