Reusing multi-scale detector features within a modified multi-object tracking pipeline reduced tracker-side processing times to between 6.83 and 11.38 milliseconds per frame on a highway-surveillance sequence, according to research published in Frontiers in Computer Science. The investigation, conducted by Ayoub El-Alami, Younes Nadir, and Khalifa Mansouri from Université Hassan II de Casablanca, evaluated how eliminating dedicated appearance networks affects the balance between computational speed and tracking consistency during vehicle tracking.
In conventional tracking systems, deploying separate deep appearance encoders to extract object representations imposes substantial computational overhead. To mitigate this computational cost, the authors modified the DeepSORT pipeline to directly repurpose neck features extracted from a multi-scale YOLOv5s object detector. Under this framework, raw bounding box predictions from the detector are mapped directly to detector-derived appearance descriptors, which are subsequently clustered using spatial overlap to generate vehicle representations without invoking an independent feature extraction model.
To regulate memory usage and representation complexity across varied object detections, the proposed architecture introduces a retention parameter designated as L. This parameter enforces a strict numerical limit on the number of retained feature vectors stored per dimensionality according to confidence rankings. When associating candidate detections with existing target trajectories, appearance matching costs are computed using the minimum cosine distance between compatible descriptors. The researchers tested three specific parameter settings for L, examining configurations with values of 5, 10, and 20.
The experimental evaluation assessed the tracking configurations on a single highway-surveillance video sequence under controlled hardware limits capped at four central processing unit cores and 4 gigabytes of random-access memory. The proposed feature-reuse pipelines were benchmarked directly against five established tracking frameworks: standard DeepSORT, baseline SORT, StrongSORT, OC-SORT, and ByteTrack. Across all evaluations, the timing metrics focused exclusively on tracker-side processing, explicitly excluding the execution time required for object detector inference and video decoding operations.
The resulting tracker-side execution speeds of 6.83 to 11.38 milliseconds per frame corresponded to approximately 15.2 percent to 25.3 percent of the processing time measured for the standard DeepSORT algorithm under identical experimental conditions. These measurements demonstrated that recycling detector-derived feature representations substantially accelerates the isolated tracking stage by bypassing the computational demands of a standalone appearance model. However, this efficiency gain occurred alongside measurable differences in identity association consistency when compared to full appearance-tracking architectures.
Tracking consistency was quantified by measuring association disagreement against a visually reviewed reference trajectory initialized with DeepSORT. For the proposed configuration using an L value of 20, association disagreement measured 0.43 percent under baseline conditions with zero detection dropout. When synthetic frame-level detection dropouts were introduced, disagreement increased to 0.96 percent at a 25 percent dropout rate and reached 1.69 percent at a 50 percent frame-level dropout rate. Throughout these trials, standard DeepSORT and StrongSORT achieved lower disagreement rates than the proposed feature-reuse framework.
The testing protocol relied on several specific operational conditions, including scoring only frames containing available detections, applying a single random pattern per nonzero dropout rate, and suspending track deletion during dropout periods. The authors noted that their experimental observations describe a distinct trade-off between computational processing cost and association agreement under this specific testing protocol. Crucially, the researchers emphasized that standard multi-object tracking metrics, performance generalization across standard public benchmarks, and end-to-end real-time execution speeds were not evaluated in the study.
Subject of Research: Computational efficiency and association accuracy in vehicle tracking through detector feature reuse
Article Title: Improving vehicle tracking efficiency through confidence-ranked YOLO feature reuse
Article References: El-Alami, A., Nadir, Y., & Mansouri, K. (2026). Improving vehicle tracking efficiency through confidence-ranked YOLO feature reuse. Frontiers in Computer Science, 8, Article 2011972. https://doi.org/10.3389/fcomp.2026.2011972
Image Credits: AI Generated
DOI: 10.3389/fcomp.2026.2011972
Keywords: multi-object tracking, DeepSORT, YOLOv5s, feature reuse, appearance descriptors, association disagreement, SORT, StrongSORT, OC-SORT, ByteTrack, cosine distance, highway surveillance, detection dropout, spatial overlap, vehicle tracking, computer vision, tracker runtime, hardware constraints
News Source: Denise Maddox. (November 18, 2026). Reusing YOLO Features Cuts Tracking Time but Increases Vehicle Association Disagreement. Scienmag.



