Stability–Plasticity Trade-Off in Sequential Adaptation of Traffic Sign Detectors Under Environmental and Cross-Dataset Shifts

Traffic sign detectors deployed in changing driving environments must adapt to newly encountered conditions while retaining previously acquired detection knowledge. This study investigates the stability–plasticity trade-off in sequential adaptation of YOLO-TS across environmental and cross-dataset shifts using TT100K, Generated-TT100K-weather, and CCTSDB2021. Four sequential adaptation scenarios were evaluated by comparing full-network adaptation (freeze = 0) with highly constrained head-focused adaptation (freeze = 17), in which 96.23% of the model parameters were frozen. Knowledge retention and new-domain adaptation were quantified using backward transfer (BWT) and adaptation gain (AG), respectively. In the cross-dataset transitions to CCTSDB2021, head-focused adaptation improved BWT from −83.21 to −3.16 percentage points (pp) and from −75.65 to −2.27 pp while maintaining AG values of +57.36 and +74.26 pp, respectively. In contrast, for the TT100K-to-Generated-TT100K-weather transition, full-network adaptation achieved a larger AG (+38.07 vs. +21.45 pp) with comparable BWT (−4.67 vs. −5.05 pp). These results demonstrate that parameter freezing does not provide a uniform advantage across sequential transitions. Instead, the appropriate scope of parameter updating depends on the characteristics of the encountered domain shift, highlighting the need to balance adaptation and retention during sequential adaptation of traffic sign detectors.