Improved DeepSORT pedestrian tracking using lightweight horizontal-stripe features
Abstract
Partial occlusion in low-altitude UAV imagery degrades pedestrian appearance cues, motion estimates, and target identity continuity. We propose an improved DeepSORT framework combining a global-local horizontal-stripe re-identification network (GLHS-ReID), an adaptive Kalman filter, and long-term appearance-memory-based rebinding. GLHS-ReID fuses a global descriptor with four stripe descriptors into a 1,280-dimensional embedding. The adaptive filter adjusts process and observation noise using detection confidence, center displacement, and velocity residuals, while long-term memory supports target recovery after track deletion. GLHS-ReID improves mAP and Rank-1 by 3.10 and 4.80 points on Market 1501 and by 2.90 and 4.20 points on Occluded-REID. The adaptive filter reduces center jitter by 10.21-19.35%. In short term occlusion, rebinding reduces reacquisition time from 1.86 to 1.21 s and improves target-retention success from 87.14% to 92.86%.