Optimization of Lane Detection by Feature Fusion Layer and 3D Object Detection by using Attention Modules

Abstract

Advanced Driving Assistance System (ADAS) are electronic systems that assist drivers in driving and parking functions. ADAS systems use automated technology, such as sensors and cameras, to detect nearby obstacles or driver errors, and respond accordingly. In this area, lane detection and 3D object detection can help the driver judge the lanes and vehicles and drive or turn. So lane detection and 3D object detection is an essential part of automatic driving. Modern cars incorporate an increasing number of driver-assist features in the lane detection research, among which automatic lane keeping. Traditional lane detection methods are used to process the image. Some recent approaches use deep learning models that have been trained to segment pixel-oriented lanes and increase the reception area. For accuracy, False Positive (FP) and False Negative (FN) were used to evaluate the TuSimple dataset. False Positive (FP) is negative cases divided into positive ones, and False Negative (FN) is positive cases divided into negative ones. Point Instance Network (PINet) based on the key points estimation and instance segmentation approach can get the highest accuracy on the TuSimple dataset. So far, Point Instance Network (PINet) has achieved 96.75% accuracy with the False Positive (FP) value of 0.0310 and the False Negative (FN) value of 0.0250 on the TuSimple dataset. After the lane detection results reach saturation, False Positive (FP) influence on the detection results is significant. The detection of the wrong lanes will cause the wrong operation in automatic driving. It is very dangerous. To improve the rationality of lane detection, we propose a dislocation fusion layer (DFL) architecture. This architecture includes four stacked hourglass networks that are trained simultaneously. It can be trained regardless of the number of traffic lines. The dislocation fusion layer (DFL) achieved an accuracy of 96.54% with the False Positive (FP) value of 0.0246 and the False Negative (FN) value of 0.0301 on the TuSimple dataset. Our results yield the minimum False Positive (FP) value on the TuSimple dataset. In the 3D object detection research, 3D sensor Lidar can capture three-dimensional objects, such as vehicles, cycles, pedestrians, and other objects on the road. Although Lidar can get the point clouds in 3D space, Lidar still lacks the fine-resolution of 2D information. Therefore, Lidar and camera fusion has gradually become a practical method for 3D object detection. Previous strategies focused on the extraction of voxel points and the fusion of feature maps. However, the biggest challenge is in extracting enough edge information to detect small objects. To solve this problem, we found that attention modules are beneficial in detecting small objects. In this work, we have developed a Frustum ConvNet and attention modules for the fusion of images from a camera and point clouds from a Lidar. Multi-layer perceptron (MLP) and the tanh activation functions are used in the attention modules. Compared with a well-known previous method, Frustum ConvNet, our method achieves competitive results, with an improvement of 0.14%, 0.31%, and 0.13% in Average Precision (AP) for 3D object detection in easy, moderate, and hard cases, respectively, and an improvement of 0.29%, 0.37% and 0.26% in AP for Bird's Eye View (BEV) object detection in easy, moderate, and hard cases, respectively, on KITTI detection benchmarks.