perception-traffic-light-recognition

Introduction

The purpose of the traffic light recognition module is to recognize the color of the traffic light. This task is completed using a conventional convolutional neural network. The input of the recognition module is an image with ROI information and a set of bounding box information as input data, and the output is a four-dimensional vector. , respectively represent the probability that each bounding box is black, red, yellow and green, if and only if the probability is large enough, the class with the highest probability will be identified as the state of the signal light. Otherwise the semaphore state is set to Unknown, indicating that the state is undetermined.

Directory Structure

├── traffic_light_recognition // trafficlight recognition module
    ├── conf            // module configuration files
    ├── dag             // dag files
    ├── data            // model params
    ├── detection       // main part for recogn
    ├── interface       // function interface folder
    ├── launch          // launch files
    ├── proto           // proto files
    ├── traffic_light_recognition_component.cc // component interface
    ├── traffic_light_recognition_component.h
    ├── cyberfile.xml   // package management profile
    ├── README.md
    ├── README_cn.md
    └── BUILD

Modules

TrafficLightRecognComponent

apollo::perception::onboard::TrafficLightRecognComponent

Input

Name	Type	Description	Input channal
`frame`	`apollo::perception::onboard::TrafficDetectMessage`	trafficlight message	/perception/inner/Retection

Note: The input channel is structure type data. The default trigger channel is /perception/inner/Retection. The detailed input channel information is in modules/perception/traffic_light_recognition/dag/traffic_light_recognition.dag file. By default, the upstream components of the messages received by the component include traffic_light_detection.

Output

Name	Type	Description	Output channal
`frame`	`apollo::perception::onboard::TrafficDetectMessage`	trafficlight message	/perception/inner/Tracking

Note: The output channel is structure type data. The message is defined in the modules/perception/common/onboard/inner_component_messages/traffic_inner_component_messages.h file. The output channel message data can be subscribed by components in the same process. The detailed output channel information is in modules/perception/traffic_light_recognition/conf/traffic_light_recognition_config.pb.txt file.

configuration

component configuration files： modules/perception/traffic_light_recognition/conf/traffic_light_recognition_config.pb.txt

parameter type	parameter name	default value	meaning
string	PluginParam.name	EfficientNetRecognition	detection algorithm name
string	PluginParam.config_path	perception/traffic_light_recognition/data	configuration file path
string	PluginParam.config_file	efficient_net.pb.txt	profile name
string	detection_output_channel_name	/perception/inner/Tracking	detection result output channel
int32	gpu_id	0	gpu id

efficient_net model configuration file location：modules/perception/traffic_light_recognition/data/efficient_net.pb.txt

parameter type	parameter name	default value	meaning
string	name	efficient_net	model name
string	framework	Onnx -> TensorRT	model inference framework
string	ModelFile.proto_file	efficient_net.onnx	model network structure
string	ModelBlob.inputs	inputs [1,3,96,96]	model input data name and dimension
int32	ModelBlob.outputs	outputs_cls、outputs_status	model output data name and dimension
int32	max_batch_size	3	Multi-batch inference input model detection dimension
float	classify_threshold	0.5	Unknown classification threshold
int32	classify_resize_width	96	model input image resize width
int32	classify_resize_height	96	model input image resize height

How to Launch

Add vehicle parameter configuration file to modules/perception/data/params in profile, corresponding frame_id and sensor_name, launch transform

bash

cyber_launch start /apollo/modules/transform/launch/static_transform.launch

Launch trafficlight perception

bash

cyber_launch start /apollo/modules/perception/launch/perception_trafficlight.launch