Deep Learning Model for Handling Environmental Variability and Tracking Occluded Vehicles for Video Surveillance
TL;DR
An adaptive occlusion-aware multi vehicle tracking model that improves robustness under varying illumination, congestion, and occlusion conditions while maintaining efficient performance is presented.
Abstract
Video Surveillance for Vehicle Detection (VD) and Tracking has many challenges like low brightness, low contrast, noise, occlusions, and identity consistency. This paper presents an adaptive occlusion-aware multi vehicle tracking model that improves robustness under varying illumination, congestion, and occlusion conditions while maintaining efficient performance. The video frame is enhanced using LAB–CLAHE and Contrast Enhancement. The proposed work incorporates an Occlusion Detection and Adaptive Handling module that analyzes motion characteristics and spatial overlap before VD. Occlusion status is recorded for each affected vehicle to support subsequent tracking and performance evaluation. This adaptive mechanism enhances the robustness of the VD by maintaining reliable vehicle localization and improving detection continuity in crowded traffic scenes with frequent object overlap. The proposed model provide the highest metric values of mAP@50 as 98.1%, mAP@50-95 as 71.5%, Precision as 92.1%, Recall as 94%, F1 score as 90% for VD, and MOTA as 85.2%, IDF1 score as 92.6% for vehicle tracking.