Deep learning based 3D target detection for indoor scenes

Instance segmentation is a challenging task that requires both instance-level and pixel-level prediction and it has a wide range of applications in autonomous driving, video analysis, scene understandingand so on. The currently dominant instance segmentation methods have excellent accuracy, but they are slow, and the processing speed will be even less satisfactory if the input is a large-scale image. In order to improve the efficiency and accuracy of instance segmentation of large-scale images, this article modifies the backbone network based on YOLACT network, adds a multi-information fusion module and provides an improved BiFPN method to achieve multi-scale feature fusion, while adding two branches to the first level detector Reti-naNet to achieve instance segmentation. The network model is tested on Cityscapes dataset and the results of the experiments show that the improved instance segmentation network in this article improves the accuracy while ensuring the speed of segmentation. The optimized network model size was reduced by 17% compared to YOLACT, and the mAP, mAP50, and mAP75 were improved by 18.3%, 32.1%, and 24.6%, respectively.

Section: Loss Functionmentioning

confidence: 99%

Section: Large Scale Instance Segmentationmentioning

confidence: 99%

Large scale instance segmentation of outdoor environment based on improved YOLACT

Zhao

Tong

Sun

et al. 2022

“…Currently, the common improvement methods are to improve on the mainstream models and design them specifically for the feature characteristics of small targets. Based on the difference of ideas, the approaches can be broadly classified into multi-scale feature prediction, [39][40][41][42][43] improving feature resolution, 44,45 extracting contextual information, 46,47 designing backbone networks and training strategies. [48][49][50][51] Deep learning has been shown to have excellent performance, [52][53][54][55][56][57][58] and an increasing number of researchers have launched studies on target detection based on this.…”

Section: Related Workmentioning

confidence: 99%

Improved single shot detection using DenseNet for tiny target detection

Wang

Sun

et al. 2022

As the development of deep learning and the continuous improvement of computing power, as well as the needs of social production, target detection has become a research hotspot in recent years. However, target detection algorithm has the problem that it is more sensitive to large targets and does not consider the feature-feature interrelationship, which leads to a high false detection or missed detection rate of small targets. An small target detection method (C-SSD) based on improved SSD is proposed, that replaces the backbone network VGG-16 of the SSD network with the improved dense convolution network (C-DenseNet) network to achieves further feature fusion through fast connections between dense blocks. The Introduction of residuals in the prediction layer and DIoU-NMS further improves the detection accuracy. Experimental results demonstrate that C-SSD outperforms other networks at three different image scales and achieves the best performance of 83. A 8% accuracy on the PASCAL VOC2007 test set, proving the effectiveness of the algorithm. C-SSD achieves a better balance of speed and accuracy, showing excellent performance in rapid detection of small targets.

“…However, the calibration of the fiber-optic stereo compound eye is a nonlinear mapping relationship, so it is difficult to establish an accurate geometric model. Aiming at this problem, this article introduces neural network, [37][38][39][40][41][42][43] which has strong nonlinear mapping ability and robustness, and the computation principle [44][45][46][47][48][49][50] of the optic stereo compound eye is studied in the next part.…”

Section: Calibration Principlementioning

confidence: 99%

Development of neural‐network‐based stereo bionic compound eyes with fiber bundles

Chen

Zhao

et al. 2022

The compound eye system has many unique advantages, enabling organisms to quickly and accurately obtain the three-dimensional spatial information of the target. Therefore, its bionic applications have great potential in object localization, dynamic tracking, and 3D reconstruction. However, the current bionic compound eye systems still have great difficulties in target detection. Most of the bionic compound eye systems only verify the possibility of three-dimensional (3D) detection, and there is no suitable calibration and detection scheme. In this article, a fiber-optic stereo bionic compound eye is designed, and a simple and rapid calibration method suitable for this system is selected by comparing two effective optical information. Further, the system quantitatively analyzes the detection performance of the target to achieve precise positioning and dynamic trajectory tracking. In the static detection experiment, the average detection errors of the 3D position at the horizontal angle, the elevation angle and the depth are 0.33 • , 0.32 • and 1.84 mm, respectively. In the dynamic trajectory tracking experiment, the circle radius errors in the X-Z plane and Y-Z plane are 0.04 and 2.20 mm, respectively. The proposed system provides a new perspective to understand the nature of the compound eye. K E Y W O R D Sbionic compound eyes 1 , dynamic track tracking 4 , neural network 5 , stereo vision system 2 , three-dimensional detection 3 INTRODUCTIONAs a natural multi-aperture optical system, due to its special physiological structure, 1,2 the compound eyes have the characteristics of small size, high integration, large field of view, 3 high resolution rate, and sensitivity to moving objects. 4 So, it has great advantages in three-dimensional (3D) target detection. The bionic compound eyes are mainly divided into planar compound eye 5-13 and curved compound eye. [14][15][16][17][18][19][20][21] The planar compound eye is small in size and has a fast response, but is limited by the range of the field of view (FOV). The curved compound eye is the closest to the biological compound eye in terms of structure and performance, and it is also the main design idea at present. However, it is difficult to acquire accurate and distortion-free optical images in a large field of view. The nature is very mysterious, the computation principle [22][23][24][25][26][27] of the compound eye is an interesting topic.Research on object detection from existing bionic compound eye shows that such a vision system has the potential for at least qualitative object detection. Richard Hornsey's team proposed a DragonflEye system based on fiber bundle coupling, and realized the tracking of planar target through vanishing point and Z-translation image center calibration techniques. 18,19 The method has a complex structure and is difficult to manufacture, though it does not require calibration of a single sub-eye and has high detection accuracy. Song et al. explored the possibility of orientation and dis-