DeltaTrack: Flow-Driven Multiple Object Tracking Accelerator with Variable LSB Approximation for Real-Time and Energy-Efficient Video Analytics

Multiple object tracking (MOT) has become a critical task in real-time video analytics, but conventional detector–tracker pipelines require executing heavy detection and tracking models at every frame, incurring high computational cost and energy consumption. In this work, we propose Delta-Track, a hardware-efficient MOT accelerator that significantly reduces redundant computations by dynamically skipping detection in non-key frames. Our method employs a lightweight optical flow estimation module to predict object trajectories in non-key frames, invoking the full detector–tracker pipeline only when new objects are detected. In addition, for hardware efficiency, we introduce a variable LSB approximation scheme, performing multiplication with reduced bit-widths in selected layers: 4-bit weights × 8-bit activations for accuracy-tolerant layers to improve performance, and 6-bit weights × 8-bit activations for accuracy-critical layers to reduce dynamic power instead of improving latency. Based on post-layout estimation in 28-nm CMOS, DeltaTrack sustains 18.5 frame/s at 640×640 (7.59 Mpixel/s) and 5.29 mJ/frame (12.9 nJ/pixel); in normalized terms, this corresponds to 2.26−4.66× higher throughput and 1.29−8.7× lower energy than prior accelerators.