Skip to content

traffic_light_classifier#

There's an implementation of CNN and HSV filter versions.

label support#

By default, this package only support typical type of traffic light in Japan. so only basic type of lamp is supported. please refer LampState

label names are assumed to be comma-separated to represent each lamp.\ For example, the traffic light with red and right arrow lit are represented by label such as "red,right", which are converted to the string array ["red", "right"] in cnn_classifier, and finally array is converted to LampState array.

HSV filter version#

CNN version#

If you use CNN, the trained file are automatically downloaded when you build.

other country traffic light model#

Please download from below list and replace onnx and label file in the data directory.

pretrained onnx model corresponding label file
United States LISA go-stop-warning-label
Japan nishishinjuku nishishinjuku-label

Reference#

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov and L. Chen, "MobileNetV2: Inverted Residuals and Linear Bottlenecks," 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, 2018, pp. 4510-4520, doi: 10.1109/CVPR.2018.00474.