An Improved Approach for Real-Time Taillight Intention Detection by Intelligent Vehicles
Open Access
Abstract
Vehicle taillight intention detection is an important application for perception and decision making by intelligent vehicles. However, effectively improving detection precision with sufficient real-time performance is a critical issue in practical applications. In this study, a vision-based improved lightweight approach focusing on small object detection with a multi-scale strategy is proposed to achieve application-oriented real-time vehicle taillight intention detection. The proposed real-time detection model is designed based on YOLOv4-tiny, and a spatial pyramid pooling fast (SPPF) module is employed to enrich the output layer features. An additional detection scale is added to expand the receptive field corresponding to small objects. Meanwhile, a path aggregation network (PANet) is used to improve the feature resolution of small objects by constructing a feature pyramid with connections between feature layers. An expanded dataset based on the BDD100K dataset is established to verify the performance of the proposed method. Experimental results on the expanded dataset reveal that the proposed method can increase the average precision (AP) of vehicle, brake, left-turn, and right-turn signals by 1.81, 15.16, 40.04, and 41.53%, respectively. The mean average precision (mAP) can be improved by 24.63% (from 62.20% to 86.83%) at over 70 frames per second (FPS), proving that the proposed method can effectively improve detection precision with good real-time performance.Keywords
Funding Information
- National Key Research and Development Program of China (2018YFB0105205, 2019AAA025, 52102399, 2022IVA039)
- Hubei Technological Innovation Special Fund (2018YFB0105205, 2019AAA025, 52102399, 2022IVA039)
- National Natural Science Foundation of China (2018YFB0105205, 2019AAA025, 52102399, 2022IVA039)
- Fundamental Research Funds for the Central Universities (2018YFB0105205, 2019AAA025, 52102399, 2022IVA039)
This publication has 38 references indexed in Scilit:
- SSD: Single Shot MultiBox Detector2016
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksIEEE Transactions on Pattern Analysis and Machine Intelligence, 2016
- Fast R-CNN2015
- Daytime Preceding Vehicle Brake Light Detection Using Monocular VisionIEEE Sensors Journal, 2015
- A Vision-Based Hierarchical Framework for Autonomous Front-Vehicle Taillights Detection and Signal Recognition2015
- Spatial Pyramid Pooling in Deep Convolutional Networks for Visual RecognitionIEEE Transactions on Pattern Analysis and Machine Intelligence, 2015
- Histograms of Oriented Gradients for Human Detection2005
- Distinctive Image Features from Scale-Invariant KeypointsInternational Journal of Computer Vision, 2004
- Performance evaluation of texture measures with classification based on Kullback discrimination of distributions2002
- A comparative study of texture measures with classification based on featured distributionsPattern Recognition, 1996