Figures (6)  Tables (5)

    • Figure 1. 

      Sensor deployment locations and hardware configurations.

    • Figure 2. 

      Framework for multi-sensor fusion perception.

    • Figure 3. 

      Network overview for the YOLOv5.

    • Figure 4. 

      Network overview for the PointPillars.

    • Figure 5. 

      Digital twin view. (a) BEV fusion view. (b) Camera view. (c) LiDAR view.

    • Figure 6. 

      Local details of the digital twin view. (a) 3D Detection Failure: The 3D detector fails to identify a vehicle near the intersection. (b) 2D Detection Success: The 2D detector successfully captures the vehicle that was missed in the 3D detection (as shown in a). (c) 3D Detection Success: The 3D detector successfully detects a vehicle that was occluded and missed by the 2D detector (as shown in d). (d) 2D Detection Failure: The 2D detector fails to detect a vehicle on the far side of the road due to the overlap of two vehicles.

    • Type Advantage Disadvantage Principle Main application Deployment location
      Geomagnetic coil[12] Fast response time, cost-effective Difficult to maintain, accuracy sensitive Electromagnetic induction Detecting the presence, passage of vehicles Position 0.2-0.4 m deep from the road surface
      Infrared sensors[13] High precision Low resolution, short distance Infrared reflectance Night vision, infrared imaging Signal arm, building above 3 m position
      Fisheye camera[14] Large detection range, high picture quality Highly expensive with distortion problems Image recognition technology Safety monitoring Signal arm, street lamp post 3−5 m height position
      LiDAR[15] 3D information available, high accuracy Slow processing, high costs Laser beam reflection Depth information perception, 3D reconstruction Signal arm, street lamp post 1.5−2.5 m height position
      Camera[5] Wide range, rich textures, low prices Light-sensitive, blurriness under high-speed Image recognition technology Path and object recognition Signal arm, street lamp post 2.5−5 m height position

      Table 1. 

      Common sensors in road infrastructure sensing systems.

    • Sensor Parameters Default Description
      LiDAR Channels 64 Number of lasers
      Height 3.17 m Height with respect to the road surface
      Range 100 m Maximum distance to measure/ray-cast in meters
      Rotation frequency 20 Hz LiDAR rotation frequency
      Points per second 500,000 Number of points
      Upper FOV 5 Angle in degrees of the highest laser beam
      Lower FOV −35 Angle in degrees of the lowest laser beam
      Noise stddev 0.01 Standard deviation of the noise model of point
      Dropoff rate 20% General proportion of points that are randomly dropped
      Dropoff intensity limit 0.8 Threshold of intensity value for exempting dropoff
      Camera FOV 90 Angle in degrees
      Focal length 360 Optical characteristics of camera lenses
      Principal point coordinate (320, 240) Image center coordinates
      Resolution 640 × 480 Measure of image sharpness
      FOV indicates field of view; stddev indicates standard deviation.

      Table 2. 

      Parameter configuration and description for Camera and LiDARs.

    • Model Test dataset mAP FPS
      YOLOv5[24] COCO 50.4% 140
      YOLOv4[25] COCO 48.9% 120
      SSD[26] COCO 41.2% 59
      Faster R-CNN[26] COCO 42.7% 7
      mAP indicates mean average precision. FPS indicates frames per second.

      Table 3. 

      Comparison of common 2D detection algorithms average performance for cars.

    • Model Test dataset mAP FPS
      VoxelNet[28] KITTI dataset 65.11% 30
      SECOND[29] KITTI dataset 76.48% 20
      Pointpillars[27] KITTI dataset 74.99% 62
      F-PointNet[30] KITTI dataset 70.39% 5.9
      mAP indicates mean average precision. FPS indicates frames per second.

      Table 4. 

      Comparison of common 3D detection algorithms average performance for cars.

    • Parameter Description Value
      PointPillars YOLOv5
      Range Detection range of the model. [0, −39.68, −3, 69.12, 39.68, 1] −
      Voxel size Voxel is a pixel in 3D space, voxel_size represents the size of the voxel. [0.16,0.16,4] −
      No. of classes Class of detection objects. 1 1
      Lr Learning rate, which determines the step size of parameter updates during the optimization process. 0.003 0.01
      Batch size Refers to the number of samples entered at once when training the model. 4 16
      Epoch Number of iterations during training. 80 300

      Table 5. 

      Hyper-parameter configuration.