Research on Target Detection and Feasible Region Segmentation Based on Deep Learning

Likai Li; Chihua Lu; Bin Zou

doi:10.3788/LOP57.121013

Laser & Optoelectronics Progress, Volume. 57, Issue 12, 121013(2020)

Research on Target Detection and Feasible Region Segmentation Based on Deep Learning

Likai Li¹, Chihua Lu^1,2, and Bin Zou^1,2、*

¹Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan, Hubei 430070, China

²Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan, Hubei 430070, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Structure of ResNet

Download full size

Fig. 2. Network structure of joint method for image detection and segmentation

Download full size

Fig. 3. Feature prediction maps of different scales; (a) 20×15; (b) 10×8

Download full size

Fig. 4. Prior box for SSD. (a) Image with ground truth box; (b) feature image with a size of 10×8

Download full size

Fig. 5. Flow chart of multi-task joint method

Download full size

Fig. 6. Graph of learning rate and loss function. (a) Learning rate; (b) loss function

Download full size

Fig. 7. Detection and segmentation of data in the joint method. (a) Scene A; (b) scene B

Download full size

Fig. 8. Actual effect figure of intelligent vehicle joint method. (a) Scene A; (b) scene B

Download full size

Table 1. Structure diagram of ResNet-50 network

View table

Table 1. Structure diagram of ResNet-50 network

Layer	Output size	Conv layer
Pre-block	160×120	$\begin{matrix} [7 \times 7,64] \end{matrix}$
Block1	80×60	$\begin{matrix} [\begin{matrix} 1 \times 1 \\ 3 \times 3 \\ 1 \times 1 \end{matrix} \begin{matrix} , & 64 \\ , & 64 \\ , & 256 \end{matrix}] \end{matrix}$ ×3 Conv
Block2	40×30	$\begin{matrix} [\begin{matrix} 1 \times 1 \\ 3 \times 3 \\ 1 \times 1 \end{matrix} \begin{matrix} , & 128 \\ , & 128 \\ , & 512 \end{matrix}] \end{matrix}$ ×4 Conv
Block3	20×15	$\begin{matrix} [\begin{matrix} 1 \times 1 \\ 3 \times 3 \\ 1 \times 1 \end{matrix} \begin{matrix} , & 256 \\ , & 256 \\ , & 1024 \end{matrix}] \end{matrix}$ ×6 Conv
Block4	10×8	$\begin{matrix} [\begin{matrix} 1 \times 1 \\ 3 \times 3 \\ 1 \times 1 \end{matrix} \begin{matrix} , & 512 \\ , & 512 \\ , & 2048 \end{matrix}] \end{matrix}$ ×3 Conv

Table 2. Detection ability of different methods
View table
Table 2. Detection ability of different methods
Detection model Feature network mAP /% FPS /frame Dataset
Faster R-CNN VGG16 89.23 7 cityscapes
Faster R-CNN ResNet-101 91.80 8 cityscapes
YoLo v2 GooleNet 83.46 21 cityscapes
Ours ResNet-50 90.15 32 mixed dataset

Table 3. Segmentation ability of different methods
View table
Table 3. Segmentation ability of different methods
Split model MIoU FPS /frame Dataset
PSPNet 76.6 2 cityscapes
ResNet-38 86.2 5 cityscapes
DeepLab v3 77.8 8 cityscapes
Ours 87.6 32 mixed dataset

Tools

Get Citation

Copy Citation Text

Likai Li, Chihua Lu, Bin Zou. Research on Target Detection and Feasible Region Segmentation Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2020, 57(12): 121013

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites