VEDA-Net: A Lightweight CNN Framework with Differential Attention and Edge-Aware Fusion for Real-Time Vehicle Detection

Abstract

Vehicle detection with a camera-based vision sensor in real time settings is critical for ensuring both the dependability and safety of AVS. In this paper we propose vehicle detection architecture using differential attention mechanisms and edge aware features (VEDA-Net). This is a lightweight CNN model which uses the Edge-Aware Feature Extraction (EAFE) technique to extract edge aware features from images. Dual Path Fusion (DPF) for fusing dual paths of information; and Differential Attention Module (DAM) to enhance the detection robustness in complex traffic environment. Additionally, VEDA-Net has a Compact Head for Detection (CHD) module that outputs bounding box and class probability through an anchor free regression method. Due to its low latency characteristics, VEDA-Net is suitable for use as an edge-based Autonomous vehicle system (AVS) perception module. On the COCO traffic dataset, VEDA-Net achieved an accuracy rate of 81%, a recall rate of 79%; and mAP@0.5 of 94%. Compared to YOLOV5s; Efficient Det-D0; and Tiny-YOLOv4, VEDA-Net's performance increased by 12%; 6%; and 16% respectively. Averaging over 500 epochs at a frame rate of 30 FPS, the average inference latency was 21.28ms. Based on qualitative results, the model demonstrated high levels of robustness under various levels of occlusions; different lighting conditions; and real-world clutter. By focusing on structural feature extraction, and selectively attending to consistent spatial regions in vehicles, VEDA-Net provides high detection robustness. As such, due to its modular and compact nature, VEDA-Net will be suitable for use as a detection module within edge-based AVS applications that require rapid, accurate, and light-weight perception modules.