LUD-YOLO Based Lightweight Object Detection With Semantically Aware Re-Pathing Model for Multi-Sensor Drone-Based Indoor Object Tracking System

Abstract

Drones, also referred to as unmanned aerial vehicles (UAVs), have evolved into an effective solution for autonomous detection and data collection in indoor spaces without the need for human involvement. It was more suitable for crucial applications, but there are several impacts faced in indoor surveillance, like battery power constraints, unreliable communication, and poor maintenance of coverage across large indoor areas. In order to improve the performance of indoor surveillance, several AI-based approaches are used for detecting objects and ensuring drone pathing according to the targeted object in indoor environments with obstacle avoidance. However, these methods include the inability to accommodate very dynamic indoor environments, where they do not have the ability to find unexpected occurrences of obstacles in real-time and the lack of reliability in low light or sensor-degradation scenarios, which limit the resilience of indoor drone surveillance systems. To overcome these issues, a lightweight UAV detection–You Only Look Once based object detection with a deep reinforcement learning method was implemented for re-planning the drone's path to avoid collision. Initially, the drone was designed with no GPS for indoor applications. The drone contains two sensors, namely, stereo cameras and light detection and ranging. The stereo cameras are used for capturing images of indoor environments. The LIDAR generates a 3D point cloud of the indoor environments, and it shows the objects' distances and shapes for exact mapping. Here, the hybrid AI model was implemented for object detection and rephasing the drone. The lightweight UAV detection–You Only Look Once (LUD-YOLO [C3K2]) detects objects based on the stereo cameras' images. After that, semantically driven deep reinforcement learning (SD-DRL) was used for re-pathing the drone based on LIDAR distance measuring if the obstacles are detected in the moving path of the drone. Finally, the proportional-integral-derivative controller adjusted the drone's motion through its pitch, yaw, thrust, and roll for stable collision-free navigation in indoor environmental monitoring. The proposed LUD-YOLO (C3K2) model achieves accuracy, precision, recall, and specificity values of 97.01%, 89.72%, 84.93%, and 88.67%. Additionally, the proposed reinforcement learning-based deep deterministic policy gradient model attains an average reward comparison value of 320, a root mean squared error value of 1.2, and a success rate value of 86%. The proposed model attains accurate object detection and efficient re-pathing without collisions, ensuring reliable surveillance in indoor environments.