A DRL based 4-r Computation Model for Object Detection on RSU using LiDAR in IloT

Internet of vehicle (IoV) network comprises Road Side Unit (RSU), which has become a computation and communication device for effective LiDAR data communication (ex: object detect information) between vehicle-to-infrastructure (V2I) and vehicle-to-vehicle. However, the LiDARs generate a massive volume of 3D data with a notable redundancy rate leads to inadequate object detection accuracy, and the high operational cost of RSU due to inadequate resource and time consumption. Estimating the computation capacity for RSU selection is an NP-hard problem. To address this issue, we propose a Deep Reinforcement Learning (DRL) influenced 4-r computation model to measure RSU cost based on resource feasibility factor and object region detection rate based on novel region-of-interest (RoI) strategy. The resource feasibility factor appraises the residual capacity and cost of RSU based on a criterion of optimality. The RoI strategy eliminates irrelevant points, noise and ground points based on distance and shape measures of an object on RSU with feasible consumption of computation resources. The simulation results show that our mechanism achieves 83% average object detection accuracy rate, 81% average service rate and 17% service offloading rate than state-of-art approaches.

I. Introduction

Internet of Vehicles (IoV) orchestration is an emerging paradigm in vehicular applications for the effective design of automated vehicles, in which RSU integrated LiDARs plays a vital role to accomplish the target. LiDAR sensors are frequently utilized in autonomous vehicles based on a detailed investigation of their deployment. The Light Detecting and Ranging (LiDAR) sensors are mainly three types. First, the Airborne sensor measures environmental and atmospheric conditions, which are mounted on satellites and aircraft. Second, spaceborne sensor measures docking distance of space station, space exploration by mounting on probe robots, and third, ground-based LiDARs to measure vehicle speed and distance, which become a prominent component for AVs, UAVs, and robots. Ground-based LiDAR sensor measures the objects with reliable depth information, which plays a vital role to localize the object with effective shape characteristics. The RSUs are enriched with limited computation resources and storage space to map the execution of latency constraint applications. The RSU-LiDAR is a static deployment. It detects the objects in abnormal environments such as rain, fog, and snow, but the LiDAR Point Clouds are sparse, irregular with high variable point density which causes occlusion and 3D space nonuniform sampling. For instance, complete traffic information is measured without noise and occlusions because of no vibrations, making the accurate computation for Autonomous Vehicles, autonomous harbours, and mining sectors. The background filter is not initiated for each object detection, but when the background changes, the performance might be diminished. In this regard, an Artificial Intelligent edge computing service is praised for IoV frameworks. Leveraging the RSU capability helps to apprehend the resource requirements to meet the QoS of connected vehicles [1], [2]. The RSUs interconnects to edge servers to provide the services to vehicles in their range of coverage with cloud service backbone. Suppose the computational service request violates the server computing capacity; the service request offloads to the cloud server for effective execution. In this scenario, 3-tired architecture is required to construct for effective vehicle orchestration. Recent computation offloading strategies pay attention to optimize the latency of applications [3], but achieving joint communication between servers is still a global challenge. The heterogeneity services (like object detection, object tracking) needs attention to balance the workload among the RSUs to enhance the service quality and system performance.