Application Service Layer

The application service layer turns a user-composed DAG into deployable processor services. Each service is defined across three repository surfaces:

Surface	Role
template/services.yaml	Catalog entry shown in DAG orchestration, including generic input and output forms.
template/processor/*.yaml	Runtime image, processor family, model parameters, queue hook, and mounted files.
dependency/core/applications/*	Service-local AI implementation loaded by the selected processor family.

The catalog uses payload-form labels such as frame, bbox, text, segmentation, track, attribute, trajectory, pose, and graph. Dayu validates form compatibility between adjacent DAG nodes while leaving application semantics to the user.

Processor families

Processor	Application export	Typical use
detector_processor	Detector	Detection-only services.
detector_tracker_processor	Detector, Tracker	Detection followed by task-local tracking.
classifier_processor	Classifier	Classification of upstream bounding boxes.
roi_classifier_processor	Roi_Classifier	ROI-cached classification.
structured_processor	Structured_Processor	Services that consume predecessor outputs or return richer structured records.

Processor results use a common content envelope with service, service-specific outputs, and a compact profile.frame_count. Structured applications return only their outputs; the processor loads frames, collects the actual predecessor results for the current DAG node, validates the record shape, and adds the envelope.

Structured traffic services

The current system provides a structured traffic family and a reviewable example at config/application_dags/driving_risk_perception.dag.

Service	Input	Output
traffic-detection	[frame]	[bbox]
road-context-segmentation	[frame]	[segmentation]
traffic-signal-recognition	[bbox]	[text]
vehicle-tracking	[bbox]	[track]
vehicle-attribute-recognition	[bbox]	[attribute]
vehicle-trajectory-prediction	[segmentation, track, attribute]	[trajectory]
pedestrian-pose-estimation	[bbox]	[pose]
pedestrian-intent-recognition	[segmentation, pose]	[text]
risk-graph-generation	[segmentation, text, trajectory]	[graph]

For implementation and extension details, see Add Applications. The code-backed service catalog, backend notes, recommended review DAG, and visualization configuration are maintained in the Structured Traffic Services reference.