Deepstream Custom test 5

Custom application based on deepstream-test5 to generate a custom JSON message, plugin nvmsgconv is also modified to generate the custom schema. All work was done with docker image nvcr.io/nvidia/deepstream:5.1-21.02-triton and tested with kafka broker.

Define JSON message

Files custom_payload.h and nvdsmeta_schema.h

We will send the next message with PGIE and SGIE metadata:

• Define CustomPayloadMeta structure where the information to generate the payment will be stored.
• Add the same parameters of the newly defined structure to the NvDsEventMsgMeta structure.

nvmsgconv

File nvmsgconv.cpp

By default msgconv generates a static payload that contains a lot of junk information. The original definition contemplates 5 schemas; place, sensor, analytics, event and object. To generate our message only one schema is needed, so we will take the analytics schema as a base and modify it to our convenience.

• Modify function generate_analytics_module_object to generate message according to our structure.
• Modify function generate_schema_message to avoid generating unwanted information.

deepstream-test5

Files deepstream_test5_app_main.c and deepstream_sgie_meta.cpp

• Include custom_payload.h
• Change generate_event_msg_meta input parameter (CustomPayloadMeta instead of NvDsObjectMeta ) and update EventMsgMeta structure with information from CustomPayloadMeta.
• Modify bbox_generated_probe_after_analytics function to update CustomPayloadMeta structure with frame and object information.

By default test5 does not use the classification information (sgie's) therefore a function is created to store information in CustomPayloadMeta (see deepstream_sgie_meta.cpp) this function receives as input the classifier list of an object (classifier_meta_list) and an instance of CustomPayloadMeta.

• Before bbox_generated_probe_after_analytics declare newly created function and after getting object information call custom_parse_sgie_meta_data.
• Modify input parameter when calling generate_event_msg_meta.
• At the end of bbox_generated_probe_after_analytics free memory allocated to user_data (instance of CustomPayloadMeta).

PS: All modified code is delimited by the line //----- CUSTOM CODE ------//.

Execution

• Copy the contents of includes/ and libs/ folders into sources/includes/ and sources/libs/ folders inside the deepstream docker.
• Compile and install nvmsgconv (export cuda and run make install)
• Copy folder deepstream-custom-test5 to folder sources/apps/sample_apps/
• Compile application (export cuda and run make)
• Modify kafka server IP and topic to subscribe to in configuration file test5_config_file_src_infer.txt