A Robust Thermal Infrared Vehicle and Pedestrian Detection Method in Complex Scenes.

Yang Liu,Hailong Su,Cao Zeng,Xiaoli Li

doi:10.3390/s21041240

Abstract

In complex scenes, it is a huge challenge to accurately detect motion-blurred, tiny, and dense objects in the thermal infrared images. To solve this problem, robust thermal infrared vehicle and pedestrian detection method is proposed in this paper. An important weight parameter β is first proposed to reconstruct the loss function of the feature selective anchor-free (FSAF) module in its online feature selection process, and the FSAF module is optimized to enhance the detection performance of motion-blurred objects. The proposal of parameter β provides an effective solution to the challenge of motion-blurred object detection. Then, the optimized anchor-free branches of the FSAF module are plugged into the YOLOv3 single-shot detector and work jointly with the anchor-based branches of the YOLOv3 detector in both training and inference, which efficiently improves the detection precision of the detector for tiny and dense objects. Experimental results show that the method proposed is superior to other typical thermal infrared vehicle and pedestrian detection algorithms due to 72.2% mean average precision (mAP).

Highlights

Thermal infrared object detection technology, due to its noncontact and passive detection characteristics, overcomes the shortcomings of radar object detection technology’s poor portability and the difficulty in detecting the objects in radar blind areas
The previous traditional infrared object detection algorithms generally used the strategy of sliding windows to select the region proposal [1,2,3,4,5], and used the algorithm based on handcrafted features to extract the shape or texture features of the objects, such as scale-invariant feature transform (SIFT) [6] and histogram of oriented gradient (HOG) [7]
The optimized anchor-free branches of the feature selective anchor-free (FSAF) module are plugged into the YOLOv3 single-shot detector with a similar feature pyramid network (FPN) [20]

Summary

Introduction

Thermal infrared object detection technology, due to its noncontact and passive detection characteristics, overcomes the shortcomings of radar object detection technology’s poor portability and the difficulty in detecting the objects in radar blind areas. It has broken through the limitations of optical object detection technology, which has a poor capability to penetrate smoke and dense fog and cannot work all day, and has become a research hotspot in the fields of military reconnaissance, traffic management, and autonomous driving. The extraction algorithms based on handcrafted features are difficult to capture the higher-level semantic information of the objects, resulting in a low detection precision

Methods

Findings

Discussion

Conclusion