Abstract
The advancement of real-time object detection and tracking systems is a crucial component of contemporary video surveillance technology. However, conventional deep learning models often struggle to perform effectively in challenging environments, such as low-resolution and low-light conditions. This paper presents Heimdall, a novel object detection and tracking solution designed to overcome these limitations. By fine-tuning Faster R-CNN models with ResNet50-FPN v1, ResNet50-FPN v2 and MobileNetV3 backbones on the ExDark dataset, we achieved significant improvements in accuracy and overall detection performance compared to default pretrained configurations. Experimental evaluations conducted on a custom dataset derived from real-world videos with low quality and taken at night demonstrated precision enhancements ranging from 0.2918 (baseline) to 0.8720 (optimized). The integration of contrast limited adaptive histogram equalization (CLAHE) further improved visibility and recall in exceptionally dark scenes. Furthermore, the integration of DeepSORT for tracking, coupled with OSNet-based re-identification, has significantly enhanced the continuity of object identification during tracking. Heimdall, with its real-time architecture and intuitive graphical user interface built on PyQt5, strikes a delicate equilibrium between precision and computational effectiveness.