Research On Object Detection And Motion Tracking Methods For Mobile Robots

With the increasing maturity of computer and communication technology and the increasing richness of mobile robot application scenarios,mobile robot technology is further iterated and upgraded.In order to adapt to the more complex and changeable dynamic operating environment,mobile robots have been developing towards higher accuracy and stronger adaptability.Through its own rich perception system and intelligent behavior system,mobile robots can realize efficient behavior ability and carry out continuous observation and monitoring in places difficult for humans to reach.Therefore,mobile robots are widely used in patrol inspection,disaster rescue,intelligent logistics,unmanned distribution and other fields.Although mobile robots have many advantages such as high autonomy and strong flexibility,they also face many challenges.At present,the detection and tracking algorithms applied to mobile robots have many shortcomings.First,the current target detection and tracking algorithms are mainly based on deep learning network.Although it has the advantages of high detection accuracy and fast identification speed,the deep learning network structure is complex and the model is huge,while the computing power of the airborne equipment of mobile robots is limited.Can not meet the complex,huge data calculation.Second,in the process of detection and tracking of mobile robots,it is inevitable that the target object will be blocked,deformed,background interference,etc.,which leads to the omission or false detection of detection and tracking algorithm,resulting in task failure.To solve the above problems,this paper uses the YOLOv5s(You Only Look Once version 5 small)detection algorithm for lightweight processing.The algorithm was fused with DeepSORT(Deep Simple Online and Realtime Tracking)multi-target tracking algorithm,and transplanted to the mobile robot hardware platform for testing,verifying the robustness of the improved algorithm in practical application scenarios.The specific work of this paper is as follows:(1)The YOLOv5s target detection algorithm is lightweight.Based on YOLOv5s,a one-stage object detection algorithm based on regression method,the backbone network of the algorithm was pruned.The specific measures are as follows:the 6 × 6 convolution in the original network is replaced by 3 × 3 convolution layer for feature extraction,and Batch Normalization(BN)layer is added after the common convolution layer to converge the network,and the common convolution module and BN layer are merged together,allowing the network to have normalization processing in the process of feature extraction.In addition,Conv and C3 modules in backbone network are replaced by lighter ShuffleNet modules,and SimSPPF(Simplified Spatial Pyramid Pooling-Fast)structure is introduced,which effectively reduces the parameter size of network model.It speeds up the reasoning speed of the network model and makes it feasible to deploy it on the mobile robot platform.(2)The improved detection algorithm is integrated with the tracking algorithm.The multi-target tracking algorithm DeepSORT was optimized for missing detection and wrong detection.The original target detection network in the algorithm was replaced by the YOLOv5s detection network with faster speed and fewer parameters.According to the target position and motion information,the relationship between the frames before and after the object was used to predict the position of the target object.A series of comparative experiments are designed to verify the reliability of the improved tracking algorithm in different situations and when the target object is blocked.(3)Pedestrian detection and tracking system based on mobile robot platform.The hardware and software parameters and the system framework of the mobile robot are introduced,and the integrated algorithm is transplanted to the airborne platform.Finally,it is verified in the real environment.The analysis of the experimental results shows that the system has high stability in the traffic environment mainly composed of pedestrians and under the interference conditions such as overlap,disappearance and deformation of target objects.