Hi all,
We are currently architecting an edge-deployed video analytics service tasked with processing 8 simultaneous 1080p @ 25fps RTSP camera streams for real-time person/object detection on an NVIDIA Jetson Orin / RTX workstation.
Our pipeline uses GStreamer with hardware-accelerated H.264 decoding feeding into a shared frame queue for batched YOLOv8/TensorRT inference. However, during high-activity periods, we observe buffer bloat and progressive stream desync:
Simplified ingest worker
import cv2
import queue
class StreamIngestWorker:
def init(self, rtsp_url, frame_queue):
self.cap = cv2.VideoCapture(rtsp_url, cv2.CAP_FFMPEG)
self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 2)
self.queue = frame_queue
def capture_loop(self):
while self.cap.isOpened():
ret, frame = self.cap.read()
if not ret:
break
if not self.queue.full():
self.queue.put_nowait(frame)
else:
# Dropping stale frames introduces decoder artifacting
pass
The Bottleneck:
If we drop intermediate non-key frames (delta frames) when the inference queue fills, we experience decoding artifacts.
If we keep the queue synchronous, latency creeps up to 2.5–4.0 seconds behind live time within 10 minutes of runtime.
My Questions:
What architectural pattern do you recommend for decoupling ingest from inference without accumulating buffer delay?
Are you relying on NVIDIA DeepStream SDK plugins (nvstreammux, nvinferserver), or rolling lightweight Python/C++ multiprocessing queues with shared memory (multiprocessing.shared_memory)?
Would love to hear how others handle multi-camera concurrency at the edge!
Top comments (0)