Hao Fu

Likelihood-Field-Model-Based Vehicle Pose Estimation with Velodyne

Dynamic vehicle tracking is an important module for Autonomous Land Vehicle (ALV) navigation in outdoor environments. The key step for a successful tracker is to accurately estimate the pose of the vehicle. In this paper, we present a novel real-time vehicle pose estimation algorithm based on the likelihood field model built on the Velodyne LIDAR data. The likelihood field model is adopted to weight the particles, which represent the potential poses, drawn around the location of the target vehicle. Importance sampling which is speeded up with the Scaling Series algorithm, is then exploited to choose the best particle as the final vehicles pose. The performance of the algorithm is validated on the data collected by our own ALV in various urban environments.

Likelihood-Field-Model-Based Dynamic Vehicle Detection with Velodyne

Dynamic vehicle detection is an important module for Autonomous Land Vehicle (ALV) navigation in outdoor environments. In this paper, we present a novel dynamic vehicle detection algorithm based on the likelihood field model for an ALV equipped with a Velodyne LIDAR. An improved 2D virtual scan is utilized to detect the dynamic objects with the scan differencing operation. For every dynamic object, a vehicle is fitted with the likelihood field model, and the motion evidence and motion consistence of the fitted vehicle are exploited to classify the dynamic object into the vehicle or not. The performance of the algorithm is validated on the data collected by our ALV in various environments.

Smooth localization independent of GPS using coarse height maps

This paper presents a new map-based localization approach for autonomous ground vehicle. To build a highly precise map, state-of-the-art methods usually explore the SLAM strategy with the assumption that it must involve loop closure. However, this assumption cannot be met in many real world scenarios. In this work we propose a novel algorithm to generate a coarse height map with the imprecise GPS pose. The vehicle position is estimated by registering feature points from the prior map with the ones in the current scans. We validate the effectiveness of our algorithm by localizing our vehicle in urban environment. Results show the method does not require GPS and can overcome the weakness of sudden jumping of GPS position and finally achieve real-time decimeter-level localization.

Partial Static Objects Based Scan Registration on the Campus

Scan registration has a critical role in mapping and localization for Autonomous Ground Vehicle (AGV). This paper addresses the problem of alignment with only exploiting the common static objects instead of the whole point clouds or entire patches on campus environments. Particularly, we wish to use instances of classes including trees, street lamps and poles amongst the whole scene. The distinct advantage lies in it can cut the number of pairwise points down to a quite low level. A binary trained Support Vector Machine (SVM) is used to classify the segmented patches as foreground or background according to the extracted features at object level. The Iterative Closest Point (ICP) approach is adopted only in the foreground objects given an initial guesses with GPS. Experiments show our method is real-time and robust even when the the signal of GPS suddenly shifts or invalid in the sheltered environment.

Augmenting cascaded correlation filters with spatial–temporal saliency for visual tracking

Abstract We herein propose a novel visual tracking approach using cascaded discriminative correlation filters (DCFs). The approach consists of two stages. In the first stage, a DCF is trained with high-level convolutional features to initially estimate the location of the object. In the second stage, another DCF is trained using low-level convolutional features to refine the object location. To efficiently track the deformable or occluded objects, spatial–temporal saliency is introduced to enhance the second stage DCF. The proposed approach is tested on the VOT2015 and OTB-13 benchmark datasets. The experimental results show that our tracker achieves state-of-the-art performance and performs extremely well in tracking nonrigid, fast moving, or occluded objects.

Multi-Object Tracking with Correlation Filter for Autonomous Vehicle

Multi-object tracking is a crucial problem for autonomous vehicle. Most state-of-the-art approaches adopt the tracking-by-detection strategy, which is a two-step procedure consisting of the detection module and the tracking module. In this paper, we improve both steps. We improve the detection module by incorporating the temporal information, which is beneficial for detecting small objects. For the tracking module, we propose a novel compressed deep Convolutional Neural Network (CNN) feature based Correlation Filter tracker. By carefully integrating these two modules, the proposed multi-object tracking approach has the ability of re-identification (ReID) once the tracked object gets lost. Extensive experiments were performed on the KITTI and MOT2015 tracking benchmarks. Results indicate that our approach outperforms most state-of-the-art tracking approaches.

Estimating Initial Guess of Localization by Line Matching in Lidar Intensity Maps.

While driving in typical traffic scenes with drastic drift or sudden jump of GPS positions, the localization methods based on wrong initial positions could not select the properly overlapping data from the pre-built map to match with current data, rendering the localizations as not feasible. In this paper, we first propose to estimate an initial position by matching in the infrared reflectivity maps. The maps consists of a highly precise prior map built with offline SLAM technique and a smooth current map built with the integral over velocities. Considering the attributes of the low-texture maps, we adopt the stable, rich line segments to match. A affinity graph to measure the pairwise consistency of the candidate line matches is constructed using the local appearance, pairwise geometric attribute and is efficiently solved with a spectral technique. The initial global position is obtained by converting the structure between current position and matched lines. Experiment on the campus with GPS error of dozens of meters shows that our algorithm can provide an robust initial value with meter-level accuracy.

Learning deep compact channel features for object detection in traffic scenes

In this work, we present a new multiple channel feature called Deep Compact Channel Feature (DCCF), which generates a compact, discriminative feature representation by a pre-trained deep encoder-decoder. With the combination of DCCF and boosted decision trees, a new object detector is proposed which achieved outstanding performance on standard pedestrian dataset INRIA and Caltech. Furthermore, a large scale and challenging Chinese Traffic Sign Detection benchmark is constructed. DCCF and other related methods are evaluated on this dataset. The dataset and baselines are available online.

Estimation of Initial Position Using Line Segment Matching in Maps

While navigating in a typical traffic scene, with a drastic drift or sudden jump in its Global Positioning System (GPS) position, the localization based on such an initial position is unable to extract precise overlapping data from the prior map in order to match the current data, thus rendering the localization as unfeasible. In this paper, we first propose a new method to estimate an initial position by matching the infrared reflectivity maps. The maps consist of a highly precise prior map, built with the offline simultaneous localization and mapping (SLAM) technique, and a smooth current map, built with the integral over velocities. Considering the attributes of the maps, we first propose to exploit the stable, rich line segments to match the lidar maps. To evaluate the consistency of the candidate line pairs in both maps, we propose to adopt the local appearance, pairwise geometric attribute and structural likelihood to construct an affinity graph, as well as employ a spectral algorithm to solve the graph efficiently. The initial position is obtained according to the relationship between the vehicles current position and matched lines. Experiments on the campus with a GPS error of dozens of metres show that our algorithm can provide an accurate initial value with average longitudinal and lateral errors being 1.68m and 1.04m, respectively.