Multi-Scale Feature Extraction and Monocular Depth Estimation Algorithm for Traffic Images

Le résumé fourni par la source

Unsupervised monocular depth estimation, also known as self-supervised monocular depth estimation, predicts the depth information of each pixel in a scene from unlabelled images or videos captured by a single camera, without requiring any manually annotated depth data. This avoids the complexity and cost of using large amounts of labelled data. Unsupervised monocular depth estimation can be used to estimate the distance between a vehicle or pedestrian and an electronic police camera by analyzing the size and position of these objects in the image, allowing the police to more accurately detect traffic violations and reduce false positives and false negatives. However, factors such as the location and angle of the camera and lighting conditions can severely affect the stability of the algorithm, resulting in semantic distortions and missing details. Therefore, this paper proposes a multi-scale feature extractor that refines the disparity module and fuses it layer by layer, while incorporating skip connections at different network layers. By training on electronic police data, the δ1 value is improved, and the maximum accuracy in identifying vehicles and pedestrians and evaluating red light violations is 97% and 97.65%, respectively.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.