Research On Robot Target Search Method In Indoor Environment

In recent years,robots are appearing more and more frequently in people’s daily lives.Target search and navigation are essential research directions in robotics,but the complexity and variability of indoor scenes can bring significant challenges to target search and navigation tasks.Many search algorithms for static objects have achieved good results,but target search algorithms for dynamic scenes are still relatively rudimentary.In this paper,we propose a robot target search algorithm for indoor scenes with an office scene as the experimental scenario,search for dynamic targets as the main task,and verify and analyze it in simulation and physical platforms.This paper establishes an initial multi-level scene semantic map based on the objects in the scene.The robot acquires image data and obtains target detection results through the camera,obtains event triads(person,time,and location)by capturing chat information within We Chat,updates the multi-level scene semantic map in real-time,and records object-to-object relationship changes,object spatial probability distribution changes,and person spatio-temporal distribution changes at different time steps in the scene according to the exponentially weighted moving average.When the robot receives the task of searching dynamic targets,it can use the cost weighting function to calculate the cost between the current position and each navigable point based on the empirical values provided by the multi-level scene semantic mapping and select the optimal strategy to execute the search task.The robot repeats this process until the target object is searched or the maximum number of searches is reached.The multi-level scene semantic mapping updates the probability distribution of dynamic objects in the scene in real-time,significantly reducing the path length required to complete the task,improving the search task’s success rate,and solving the current problem of poor robot performance when searching for dynamic targets.In addition,to facilitate the deployment of deep networks for target detection on mobile robots,an improved lightweight YOLOv5 network model is proposed in this paper.The model uses the LPC module presented in this paper to aggregate multi-scale single-channel null convolution in parallel,which significantly reduces the number of parameters of the model and improves the deployment performance of the network model while ensuring the target detection performance.