Object Detection for Mobile Robot Navigation in Dynamic Indoor Environment

Abstract

The detection and recognition of objects is a very important and challenging task in computer vision, as there is an increasing interest in building an autonomous mobile system. To make a mobile service robot truly ubiquitous to complex and dynamic indoor environments, they should be able to understand the surrounding environment beyond the capability of avoiding obstacles, navigating autonomously, and building maps. It’s necessary to build a reliable and fast detection system to enhance the performance of indoor robot navigation. Several researchers have proposed different techniques for recognizing and localizing indoor objects for mobile robot navigation, however, most have low recognition rate, and some doesn’t recognize multiple objects or does not run in real-time. Mobile robots do not come with heavy computing power so the detection algorithm response time should be sufficient enough that the robots can make a decision fairly quickly. Multiple object recognition and localization for mobile robot navigation in a dynamic indoor environment are proposed in this study. The algorithm detects an object by combining deep features for recognition and local features for localization. For recognition of multiple objects, a CNN based Deep learning model that uses MobilenetV2 as a base network with a sigmoid classification layer is developed. The model generates a probability for each object in the image independently, which is used for localizing the recognized object in the image. ORB+SURF is used as a feature extractor to generate a bounding box for each object. ORB+SURF is compared with different feature detection and descriptor combination to evaluate its performance. The object detection algorithm is also integrated with ROS to evaluate the performance in a real-world scenario and in doing so the distance information is also extracted from the Laser distance sensor mounted on the top of Turtlebot. Different navigation scenarios are executed by combining the object detection algorithm and the distance information in the test environment. The proposed object detection algorithm for mobile robot navigation achieves better performance in terms of recognition rate and speed, and also it provides the distance of the object to the robot.