A deep learning approach to traffic lights: Detection, tracking, and classification
Citations
802 citations
Cites background or methods from "A deep learning approach to traffic..."
...BSTL [84] 2017 The largest traffic light detection dataset....
[...]
...In deep learning era, some well-known detectors such as Faster RCNN and SSD were applied in traffic sign/light detection tasks [83, 84, 378, 379]....
[...]
674 citations
543 citations
154 citations
129 citations
Cites background or methods from "A deep learning approach to traffic..."
...The training strategy is consistent with BSTLD....
[...]
...We conduct extensive ablation study and experiments on multiple datasets including both aerial images from DOTA [10], DIOR [11], UCAS-AOD [27], as well as natural image dataset COCO [8], scene text dataset ICDAR2015 [28], small traffic light dataset BSTLD [29] and our newly released S2TLD to illustrate the promising effects of our techniques....
[...]
...We perform extensive ablation studies and comparative experiments on multiple aerial image datasets such as DOTA, DIOR, UCAS-AOD, small traffic light dataset BSTLD and our released S2TLD, and demonstrate that our method achieves the state-of-the-art detection accuracy....
[...]
...BSTLD [29]: BSTLD contains 13,427 camera images at a resolution of 720 × 1,280 pixels and contains about 24,000 annotated small traffic lights....
[...]
...In the experiment, we divide BSTLD training set into a training set and a test set according to the ratio of 6 : 4....
[...]
References
73,978 citations
33,597 citations
"A deep learning approach to traffic..." refers methods in this paper
...In addition, we employ two max-pooling and three dropout layers [17]....
[...]
30,811 citations
28,225 citations
"A deep learning approach to traffic..." refers background or methods in this paper
...With the recent advances and performance of deep neural networks [10]–[13], significant improvements were made in several fields of machine learning and especially computer vision....
[...]
...Deep learning has been used for image classification [10], end-to-end object detection [11], pixel-precise object segmentation [13], and other applications....
[...]
27,256 citations
"A deep learning approach to traffic..." refers background or methods in this paper
...Removing the classification part from the YOLO architecture yields the following loss function: λcoord s2∑ i=0 B∑ j=0 1objij ( (xi − x̂i)2 + (yi − ŷi)2 ) + λcoord s2∑ i=0 B∑ j=0 1objij ( ( √ wi − √ ŵi) 2 + ( √ hi − √ ĥi) 2 ) + λnoobj s2∑ i=0 B∑ j=0 1noobjij (pi) 2 + s2∑ i=0 B∑ j=0 1objij (pi − p̂i)2 (1) For a detailed explanation of all terms, please refer to [11]....
[...]
...As in [11], this part of the loss function is only used if an object overlaps with the prediction....
[...]
...The detection of traffic lights is carried out by an end-to-end trained neural network [11], which we adapt to detect traffic lights as small as 3 × 10 pixels....
[...]
...Initial tests using the original YOLO architecture showed lower performance in detection when the classification part of the network was used....
[...]
...Deep learning has been used for image classification [10], end-to-end object detection [11], pixel-precise object segmentation [13], and other applications....
[...]