An Enhanced Mobile Net-Ssd Architecture For Multi Scale Object Detection In Advanced Driving Assistance System

Abstract

An optimized real-time object detection framework is developed in this study for deployment in Advanced Driver Assistance Systems (ADAS). The proposed solution adopts a compact Mobile Net-based backbone coupled with a Single Shot Detector (SSD), further enhanced through hierarchical receptive field aggregation and an attention-guided feature integration mechanism. Validation on the KITTI benchmark indicates that the proposed model achieves a detection accuracy of 82.3% mAP while maintaining a per-frame processing time of 68 milliseconds. From a computational perspective, the network architecture is highly lightweight, comprising only 4.8 million trainable parameters. This corresponds to a 95.5% reduction in model size when compared with Faster R-CNN, which requires approximately 137 million parameters. Experimental comparisons with widely adopted detection frameworks demonstrate consistent performance gains, including an improvement of 5.1 percentage points over Faster R-CNN and a 4.2 percentage point increase relative to YOLOv3s.The incorporation of attention-driven feature fusion significantly strengthens detection capability in scenarios involving visual obstruction. Specifically, pedestrian recognition accuracy improves by 18.1% compared with standard SSD implementations. In cases of partial occlusion, the proposed method retains an accuracy of 78.1%, outperforming baseline detectors that achieve 66.3%. The system processes video streams at 14.9 frames per second with stable inference latency, confirming its suitability for time-critical automotive environments. Owing to its modular construction and reduced computational demand, the proposed framework can be seamlessly deployed within existing ADAS architectures. The presented approach effectively addresses key perception challenges in autonomous driving while satisfying the strict efficiency constraints of embedded automotive platforms.