Research On Monocular Visual Perception Method Of Moving Target In Multi-robot Confrontation Scenarios

Video surveillance technology has the advantages of all-weather monitoring,rich information,intuitive and clear,and has been applied in more and more fields,such as intelligent transportation,fire warning,and people counting.However,with the rapid development of deep learning technology based on neural network,more intelligent requirements are put forward for video surveillance,such as automatic target recognition and trajectory analysis,and acquisition of target depth information.This thesis is based on the monocular visual perception of moving targets in multirobot confrontation scenarios.The main research work and achievements are as follows:(1)The latest target detection technology yolov5 algorithm is applied to the ground robot target detection task of the RMUA Artificial Intelligence Challenge.In addition,the network model is compressed and accelerated from multiple dimensions.One is to design a compact target detection network to identify ground robots.This lightweight target detection model is designed based on lightweight classification networks such as Mobile Netv2,Mobile Netv3,and Ghost Net.The second is to use the channel pruning method to perform channel pruning on the robot target detection model.The third is to perform half-precision acceleration on the algorithm model through Tensor RT.Finally,the method of combining channel pruning and half-precision acceleration was selected.On the GPU1660 TI device,the inference speed reached 205 FPS,and the m AP was 0.832 and the parameter size was 8.6M when the recall rate was 0.5.(2)The sentry robot consists of surveillance cameras at the edge of the field,and develops an algorithm for spatial positioning of the ground robot through the sentry robot.The first method is to use the field elements to solve the pose relationship between the camera coordinate system and the field coordinate system,and then solve the two-dimensional coordinates of the ground robot on the competition field according to the camera pose relationship;the second method is to directly convert the problem into a neural network.Supervised learning regression problem to solve.On the competition field of 4.48\times8.08 m,the spatial positioning accuracy based on neural network can reduce the error to 8.95 cm compared with the method of mathematical solution,which meets the needs of realtime competition in the competition.(3)Aiming at the problem that the ground robot moves quickly and leads to the lag of spatial positioning during the competition,the long short-term memory network is used to predict the movement trajectory of the ground robot,which has a good performance in the actual competition.Finally,the lightweight target detection model and trajectory prediction model are combined into a complete engineering system.The entire software system includes functions such as ground robot recognition,spatial positioning and trajectory prediction,and local area network wireless communication.