A smart-city pilot installs a battery-powered traffic camera with an embedded CPU (no GPU). The onboard detector yields vehicle bounding boxes at 30 fps; the team must add multi-object tracking so each car keeps its ID while in view. Constraints:
≤ 10 ms extra latency per frame (CPU-only)
Brief occlusions (≤ 3 frames) should be handled; long-term re-ID is unnecessary
RAM limit: 100 MB Which tracking approach BEST satisfies these requirements out of the box?
Joint Probabilistic Data Association (JPDAF) implemented with 1,000-particle filters per track
Simple Online and Realtime Tracking (SORT) using a Kalman filter and Hungarian assignment on IoU
Offline global data association formulated as a min-cost-flow across the entire sequence
Deep SORT that augments SORT with a deep CNN appearance descriptor
SORT predicts each track's state with a constant-velocity Kalman filter and performs data association with the Hungarian algorithm on an IoU cost matrix. The original study measured 260 Hz on a single CPU core and only kilobytes of memory, easily meeting a 10 ms/frame budget and 100 MB cap; its frame-to-frame design tolerates short occlusions. Deep SORT adds a CNN appearance encoder that costs additional milliseconds (often GPU-accelerated), exceeding the edge budget. A particle-filter JPDAF propagates hundreds of particles for every target and evaluates many joint hypotheses, quickly exhausting compute and memory. Min-cost-flow trackers solve a global optimization over many frames, making them unsuitable for hard real-time latency. Therefore, SORT is the most appropriate choice for this scenario.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
How does a Kalman filter work in SORT?
Open an interactive chat with Bash
What is the role of the Hungarian algorithm in SORT?
Open an interactive chat with Bash
Why is Deep SORT unsuitable for the described constraints?