A Monocular 3D Object Detection Method Based on Lightweight Feature Pyramid Structure

Abstract

The invention discloses a monocular 3D target detection method based on a light-weight feature pyramid structure, comprising: collecting RGB images of a vehicle-mounted camera; inputting the RGB images into a pre-established and trained monocular 3D target detection network, and outputting target detection Results; The monocular 3D target detection network includes: feature extraction network, detection head and post-processing module; the feature extraction network is used to extract advanced semantic features from RGB image downsampling, generate 4 times, 8 times and 16 times downsampling feature maps and Input to the detection head; the detection head is used to generate candidate key point category vectors and candidate key point pixel position index vectors based on 4 times downsampled feature maps, and generate candidate key point 3D regression based on 4 times, 8 times and 16 times downsampled feature maps The frame encoding vector, and output the candidate key point category vector and the 3D regression frame encoding vector to the post-processing module; the post-processing module is used to decode the 3D regression frame encoding vector, and output the target detection result in combination with the candidate key point category vector.

Description

technical field [0001] The invention relates to the technical field of automatic driving, in particular to a monocular 3D object detection method based on a lightweight feature pyramid structure. Background technique [0002] In the automatic driving system, 3D object detection is a very important task in the perception module. The back-end prediction, planning, motion control and other modules all rely on the reliable detection results of specific types of objects around the main vehicle. With the advantages of high-beam lidar that can accurately model the surrounding environment at the centimeter level, lidar-based 3D target detection algorithms have made great progress in recent years. Congenital deficiencies such as poor capabilities severely limit the large-scale implementation of lidar and related algorithms in the field of autonomous driving. Compared with lidar, vision sensors are not only low in cost, better than lidar in adapting to severe weather such as rain and...

AI Technical Summary

Problems solved by technology

[0004] The idea of ​​no anchor frame is the latest research hotspot in the field of target detection. Although the newly proposed pure visual 3D target detection algorithms based on key points (such as CenterNet, SMOKE, RTM3D) have high algorithm efficiency (CenterNet: 30ms, SMOKE: 30ms , RTM3D: 50ms) to meet the requirements of real-time deployment and engineering implementation on the edge computing platform for autonomous driving, but the low accuracy index still cannot fully meet the perception needs in the autonomous driving scenario

[0005] Subsequent improvement methods mainly include: adding traditional feature pyramid structure, connecting multi-stage cascaded regression structure, and introducing deep reinforcement learning to realize the optimization of detection frames. These improvement schemes have effectively improved the accuracy index of the algorithm, but due to the The additional structural branch introduced also greatly increases the delay of model reasoning, which affects the real-time deployment of the algorithm on the edge computing platform for autonomous driving; therefore, how to greatly improve the accuracy index of the existing method without reducing the efficiency will be Has great practical engineering value

Images