Model based vehicle localization for urban traffic surveillance using image gradient based matching

Abstract

The matching between 3D model projection and 2D image data is a key technique for model based localization, recognition and tracking problems. Firstly, we propose a fitness function to evaluate the matching degree that uses image gradient information in the neighborhood of model projection. The weighting adjustment and the normalization for visible model projection are involved, which improves the correctness and robustness of fitness function. The fitness function is used for vehicle localization and the 3D pose is reduced to location and orientation. Then, we present a direct search optimization method with 3×3 search kernel for location estimation. The "disturbed particles" is used to avoid falling into local optimum and the coarse-to-fine optimization strategy is adopted to greatly reduce computational cost. Finally, we propose a 3D pose estimator to find location and orientation by optimizing the fitness function within orientation range. Experiments on real traffic surveillance videos reveal that the proposed optimization algorithm is effective and both fitness function and 3D pose estimator are correct and robust against clutter and occlusion.

FCN-rLSTM: Deep Spatio-Temporal Neural Networks for Vehicle Counting in City Cameras

“…2. Detection based methods [47,36] detect individual vehicles in each frame and perform poorly in low resolution and high occlusion videos.…”

Deformable Model-Based Vehicle Tracking and Recognition Using 3-D Constrained Multiple-Kernels and Kalman Filter

“…Overview of the proposed system mate the 12 shape and three pose parameters iteratively. The approach in [35] propose an efficient search method using 3 × 3 search kernel and accelerated step length, to reduce the computational cost for the tracked poses.…”

Object Detection and Tracking Algorithms for Vehicle Counting: A Comparative Analysis

“…Vision-based vehicle counting is an interesting computer vision problem tackled by different techniques. As per the taxonomy accepted in (Zhang et al 2017), the counting approach could be broadly classified into five main categories: counting by frame-differencing (Cucchiara et al 2000; Tsai and Yeh 2013), counting by detection (Toropov et al 2015;Zheng and Peng 2012), motion based counting (Chen et al 2010;Chen et al 2012;Mo and Zhang 2010;SuganyaDevi et al 2012), counting by density estimation (Lempitsky and Zisserman 2010) and deep learning based counting (Arteta et al 2016;Hsieh et al 2017;Onoro-Rubio and López-Sastre 2016;Sindagi and Patel 2018;Zhang et al 2015Zhang et al , 2016Zhao et al 2016). The first three counting methods are environmental sensitive and generally don't perform very well in occluded environments or videos with low frame rates.…”