Small Object Detection Based on Stereo Vision

Abstract

Small size objects which dimensions are around 0.15m are one of the major security risks to driving vehicles in the highway. Lidar and radar are hard to detect this kind of objects due to the sparsity of their detecting signal. Vision based methods are possible to solve this problem because camera can generate dense information. We propose a new method to detection small objects in the highway based on stereo vision. This method uses Multi-Path-Viterbi algorithm to obtain dense depth information of stereo images. Based on the depth information, road surface can be detected. Objects on road can be mapped to the 3D space to determine their size and location, then small objects dangerous to the host Vehicle can be recognized and located.

AMMDAS: Multi-Modular Generative Masks Processing Architecture With Adaptive Wide Field-of-View Modeling Strategy

“…There are methods & papers which utilized the above mentioned hardware on bottom bases and have shown some significant contribution towards DAS systems. Some methods take the input feed and process either in 2D [3], [4], [13], [16], [5], [7], 3D point cloud [17], [6], [14], [18], [11] format or in dense disparity depth maps each of the methods implemented show both advantage & spike in productivity along with disadvantages & increased production costs for their respective implementations. Moreover, some methods concentrate only on object flow [3], [13] , [4], [9] and behavior analysis, some are biased towards a specific class [3], [4], [11] to identify proximity, and many methods analyze the correlation and occurrence between the objects present in FOV with respect to host source using 2D/3D image processing, disparity map processing techniques and some by implementing DNNs.…”

“…In Quin Long et.al. [14], Proposed a new methodology for identifying small objects in the roadway based on stereo vision. They used a multi path viterbi algorithm to obtain dense depth data of stereo images.…”

“…The above discussed papers show sensitivity towards wild, noisy conditions, and do not consider a wide field of view ( [3], [16], [13], [5], [6], [14], [9], [7], [15]). Either in some papers, the input is a stereo feed where they merge the feed for their implemented algorithm to process (to create disparity or depth maps etc) [3], [16] , [5], [6], [18], [11], [7], [15], by considering a single camera or stereo merged feed we can not capture the entire spatial surrounding in a single epoch for the algorithm to process.…”

“…These l, c, r-BOL masks are color histograms analyzed [28] to determine required parameters for the internal logic-module to frame proximity alerts and navigation suggestions. The parameters which are determined immediately after post-processing and histogram analyzing a specific [28] K-BOL mask are grouped together, as a list of several sets of pixels{PX o }, which are calculated using (14). Where each set contains a group of pixels belonging to a specific O th RGB color code; and the overall list consists of "i" number of pixel sets where i represents objects identified in that respective K-BOL mask (Fig.…”

“…In the context of driver-assistance systems [2], the purpose of obstacle-detection and tracking systems [3], [9], [13]- [15] is to detect and monitor & analyze the dynamic-behavior of one or more obstacles in the proximity of the host vehicle. The role of lane detection [12] and road spatial analysis [16], [17], [14] is to analyze and make decisions based on dynamic surroundings for safer navigation. These modules play a predominant role to dynamically process the DAS systems [2], [8] into the surrounding environment.…”