Date of Publication : April 2026

Abstract: Real time object segmentation and multi object tracking remain computationally challenging on edge devices with constrained resources (ARM processors, limited memory, low power budgets). Most of the existing solutions lean one way or the other. Vision transformers deliver strong accuracy but are far too heavy, while lightweight CNNs runs faster but sacrificing performance. We introduce EViT (Efficient Vision & Tracking), a hierarchical transformer based architecture that blends multi scale feature extraction with lightweight segmentation and Kalman filter based tracking. By using overlapping patch embeddings, efficient spatial reduced self attention, and an all-MLP decoder, EViT achieves real time performance on edge devices without compromising on the segmentation quality. Our hierarchical encoder progresses through four stages which reduces spatial resolution from 1/4 to 1/32 stride while maintaining the informative features through depthwise convolutions and multi head attention. The tracking module integrates a Kalman filter with IoU based data association for robust identity preservation across frames. Experimental validation on edge hardware (ARM Jetson Nano) shows that EViT achieves 15–20 FPS at 512 × 512 resolution with ∼17M parameters, outperforming SegFormer and MobileSAM-Track baselines. The model is deployable via TensorRT quantization (INT8) with <50 ms latency, making it suitable for real world surveillance, autonomous robotics, and constrained IoT applications.

Reference :