Vehicle target detection method based on improved YOLO V3 network model

YOLO V3 algorithm

In the aspect of target recognition, the traditional sliding window detection method has weak generalization ability and low accuracy, and the detection speed is slow. The deep learning method can effectively improve the efficiency of target detection. The RCNN target detection method proposed in 2013 can improve the speed of target detection to a certain extent through the combination of regional candidate frame and SVM classifier, but the model of this method is complex and the detection accuracy is not high. The improved Faster RCNN reduces the calculation of the model through a two-stage detection method, but it still can not meet the actual work requirements in the real-time detection. In 2016, the YOLO model was proposed to take target detection as a regression problem, which increased the test speed of target detection to 45 frames per second, and became the main method to improve the efficiency of target detection. Considering the working environment of target detection in this article, the work of this article is to improve the YOLO model.

The YOLO V3 model is the third version of the YOLO network development, and it is an important algorithm for real-time detection of targets. Based on inheriting the advantages of the previous two versions, the algorithm has made more effective improvements in multi-target detection. The improvement in the prediction of the candidate frame reduces the confidence loss of the original network model and improves the imbalance problem of object category detection (Tang, Wang & Kwong, 2017). Through the increase in the number of logical classifiers and clustering categories, the prediction effect of the target category is improved, and the feature map fusion technology of multiple scales is introduced, so that the model can be applied to target detection applications of different sizes.

The YOLO V3 model divides the acquired image into several small grid units. Its model predicts a series of candidate boxes for the location information of each grid cell and calculates its confidence. Its calculation formula is as follows: (1)${\displaystyle C_i^j=P_{ij}\left(Ob\right)\ast Io{U}_{pred}^{truth}}$ where $C_i^j$ is the confidence degree of the j-th candidate frame corresponding to the i-th unit. P_ij(Ob) is a function corresponding to the detection target. IOU${}_{pred}^{truth}$ is the intersection and union ratio of the prediction box and the real location box.

Improved YOLO V3 algorithm

The YOLO V3 algorithm can meet the image detection of multi-scale targets and can meet the needs of most image detection. However, for UAV video images, the targets of video vehicle detection are all small-sized, so it is necessary to improve the small-scale detection based on the existing algorithm model.

To make the YOLO V3 algorithm more suitable for the detection of small-scale targets, the network structure adapts to the three detection scales in the original model is optimized and improved.

The feature map of scale three is fused with the second residual block of the network. Meanwhile, the scale network suitable for large targets is deleted to improve the computational efficiency of the model and the adaptability of UAV target detection (Bruch et al., 2019). In order to make the YOLO V3 algorithm more suitable for small-scale target detection, this article optimizes and improves the network structure adapted to the three detection scales in the original model, fuses the feature map of scale three with the second residual block of the network, and deletes the scale-one network suitable for large targets, which is to improve the computational efficiency of the model and the adaptability of the UAV target detection.

In the process of using the YOLO V3 network algorithm, it refers to the idea of the anchor box, and improves the training efficiency of the network model by pre-setting the size of the candidate box. The setting of the network candidate frame is mainly obtained through K-means cluster analysis, which has a good predictive ability, but also greatly reduces the convergence speed of the model.

In response to the above problems, this article uses AvgIoU as the objective function of the model to optimize the clustering analysis method. Its expression is as follows: (2)${\displaystyle AvgIoU=\frac{\sum _{i=1}^k\sum _{j=1}^{n_k}IoU\left(b_i,c_j\right)}{n}}$ Where b_i is the i-th real label frame of the target sample, c_jrepresents the j-th cluster center of the cluster analysis algorithm, n_k is the sample size corresponding to the k-th cluster center, k is the number of clusters, and n is all the number of samples, IoU is the intersection ratio of the sample label box and the cluster center box.

In the optimized algorithm, the larger the number of clusters k is, the greater the detection stability of the objective function is, and the relationship between the objective function and the number of detection layers is considered comprehensively. Through the experiment summary, the number of anchor boxes selected by the improved algorithm is 9. The specific sizes are mainly (10,18), (8,30), (15,35), (19,55), (25,54), (24,82), (33,60), (45,88) and (57,130).

Considering that the sensitivity of different anchor boxes corresponding to different feature maps is very different, it needs to be placed on a feature map of the appropriate size. Thus, it needs to be allocated. According to the IoU between the anchor boxes and the real frame, assuming that the height and width of the anchor boxes are h and w respectively, there is a relationship between the size of the anchor boxes and the number of down-sampling d, and the distribution results of the anchor boxes are shown in Table 1.

Network optimization

For the initial YOLO v3 algorithm model, there is also the problem of deep networks. The main reason is that the size selection of the residual model is unreasonable, which can result in the stacking of multiple residual modules as well as affect the model calculation (Naqvi et al., 2022). Improving the residual module and using the combination of multiple receptive field modules can better improve the above problems. The improved residual module structure is shown in Fig. 1.

In Fig. 1, when two residual modules work on the input feature map at the same time, the convolution kernels with different receptive fields have different extraction operations on the feature map. The smaller receptive field residual module is responsible for detail feature extraction, and the larger receptive field residual module is mainly responsible for the overall feature extraction of the target. After convolution fusion, the comprehensive features of the detection target are obtained (Chen & Chen, 2020). At the same time, the low-rank decomposition method is used to decompose the 5×5 convolution kernel into two serial convolution kernels of 5×1 and 1×5, which can greatly reduce the calculation amount of model analysis under the premise of ensuring the effect of feature extraction.

The influence of the number of residual modules is considered separately. For the four scales of 104×104, 52×52, 26×26 and 13×13, the number of residual modules contained in each scale is considered to be 1∼4. After target testing, the mAP results were 78.5%, 88.7%, 82.1, and 72.7%, respectively. It can be seen from the results that the mAP values of the two residual models are the highest, which is better than the other three cases. Therefore, this article chooses to use two residual modules to complete the corresponding image detection work.

Introduction of direction angle

In the classic YOLO V3 model, the angle of the detection target is not considered. In addition, the labeling of the target is not clear and precise enough, especially when the vehicle is very close, the overlapping of the rectangular frames will lead to errors in the identification and tracking of the target. Introducing a rectangular frame with orientation angles can effectively solve the above problems (Cho & Yoon, 2018).

In the article, by positioning the six-dimensional vector ( b_x, b_y, b_w, b_h, θ, P_r) of the target position information, the angle of the target can be accurately reflected, where bx and by are the position coordinates of the target center, b_w and b_h are the width and height of the rectangular frame, θ is the direction angle of the target, and P_r is the position prediction confidence. Its expression is as follows. (3)${\displaystyle \begin{array}{c}\hfill b_x=\sigma \left(t_x\right)+c_x\hfill \\ \hfill b_y=\sigma \left(t_y\right)+c_y\hfill \\ \hfill b_w=p_w{e}^{t_w}\hfill \\ \hfill b_h=p_h{e}^{t_w}\hfill \end{array}}$

Where σ is the Sigmoid activation function of the model, p_w and p_h are the width and height of the anchor box, respectively, and (c_x, c_y) are the coordinates of the upper left corner of the anchor box.

After considering the direction angle, the loss function is also improved accordingly. Therefore, the input–output loss function and the direction-angle loss function are respectively introduced and expressed by cross-entropy. The final improved loss function can be expressed as follows. (4)${\displaystyle \begin{array}{c}L\left(b_x,b_y,b_w,b_h,{\theta }_r\right)={\lambda }_{cr}\sum _{i=1}^{s^2}\sum _{j=1}^M{W}_{ij}^{ob}\left[{\left(x_j^g-x_i\right)}^2+{\left(y_i^g-y_i\right)}^2\right]\hfill \\ +{\lambda }_{cr}\sum _{i=1}^{s^2}\sum _{j=1}^M{W}_{ij}^{ob}\left[{\left(W_{ij}^g-W_{ij}\right)}^2+{\left(h_{ij}^g-h_{ij}\right)}^2\right]\hfill \\ +{\lambda }_p\sum _{i=1}^{s^2}\sum _{j=1}^M{W}_{ij}^{ob}\left[C_{ij}^{g}log{C}_{ij}+\left(1-C_{ij}^g\right)log\left(1-C_{ij}\right)\right]\hfill \\ +{\lambda }_{cl}\sum _{i=1}^{s^2}\sum _{j=1}^M{W}_{ij}^{ob}\text{Cross Entropy}\left({\theta }_i^g,{\theta }_i\right)\hfill \end{array}}$

In the formula, s ₂ is the number of grid units; M is the number of anchor boxes; λ_cr, λ_pr and λ_cl are the weight parameters of the corresponding losses. When the coordinates of the center point in the i-th anchor box are ( $x_i^g$, y${}_i^g$) and it contains the true center point, $W_{ij}^{ob}$ = 1; otherwise, $W_{ij}^{ob}$ = 0. ${\theta }_i^g$ is the angle between the center points, and θ_i is the angle of the i-th point. For multiple objects appearing in the image, each object corresponds to a bearing angle. i is the number of targets in the detected image.

Where the expression formula of cross entropy is as follows. (5)${\displaystyle \text{Cross Entropy}\left({\theta }_i^g,{\theta }_i\right)=-\sum \underset{i}{\overset{g}{\theta }}log{\theta }_i.}$

Parameter setting

The aerial photography data set VEDAI is selected as the data set for the algorithm verification of this article. VEDAI is a dataset for vehicle detection in aerial imagery as a tool to benchmark object detection algorithms in an unrestricted environment. In addition to containing very small vehicles, the database also exhibits different variability, such as multiple orientations, lighting/shadow variations, specular reflections or occlusions. The dataset contains 1,210 aerial images of 1024×1024 with a spatial resolution of 12.5 cm. There are also multiple car targets in this data set, which is suitable for the detection and verification of small targets. The data set contains nine types of targets. They include boats, cars, campers, planes, pick-up vehicles, tractors, trucks, vans and other categories. 70% of the image data in the data set are used for model training in the article, and the remaining image data are used for testing the optimized model. The stochastic gradient descent method is selected to optimize the model, and the learning rate of the algorithm is set to 0.001, and the attenuation coefficient is 0.005.

The performance evaluation of the model is measured by the recall rate, precision and F1-score, and their expressions are as follows. (6)${\displaystyle \text{precision}=\frac{TP}{TP+FP}}$ (7)${\displaystyle \text{recall}=\frac{TP}{TP+FN}}$ (8)${\displaystyle F1=\frac{2\cdot \text{precision}\cdot \text{recall}}{\text{precision}+\text{recall}}.}$

In the formula, TP, FP, and FN are positive samples predicted as positive by the model, negative samples predicted as positive by the model, and positive samples predicted as negative by the model.

Data analysis

2. The optimization effect of the network model

Similarly, this article uses the previous VEDAI dataset, and the convergence curve of the loss function of its network model is shown in Fig. 2.

From the results in Fig. 2, it can be seen that when the number of iterations reaches 25,000, the parameters of the model are stable. The convergence of the classic YOLO V3 model is similar, with a relatively higher loss function. The VEDAI data set is used to conduct a comprehensive index test on the improved network, and the test results are compared with other algorithms. The results are shown in Table 5.

Table 5:

Comparison of detection accuracy of different indicators.

Model	Precision	Recall	F1-score	FPS
YOLO V3	73.4%	67.2%	70.2%	45
YOLO V3+RESNET34	76.5%	70.6%	71.3%	60
Algorithm after optimization of two-way residual network	88.3%	81.4%	84.7%	42

DOI: 10.7717/peerjcs.1673/table-5

As can be seen from the results in Table 5, the precision rate of the improved network model is 88.3%, which is much higher than other YOLO algorithms; and the recall rate reaches 81.4%, which is still the best result among all algorithms; the F1-score of this algorithm reaches 84.7%. However, due to the complex structure of the residual module, the algorithm is slightly lacking in reasoning speed, but it is not far behind the YOLO v3 model before improvement.

The comparison of the detection results of the improved YOLO V3 network model with orientation angle detection is shown in Table 6.

Table 6:

Comparison of detection model indicators of different algorithms.

Model	Precision	Recall	F1-score	FPS
YOLO V3	73.4%	67.2%	70.2%	45
YOLO V3-tiny (Adarsh, Rathi & Kumar, 2020b)	64.8%	59.1%	61.9%	68
Algorithm after optimization of two-way residual network	88.3%	81.4%	84.7%	42
The final optimization algorithm with improved orientation angle	89.3%	83.2%	86.1%	41

DOI: 10.7717/peerjcs.1673/table-6

From the results in Table 6, it can be seen that the detection accuracy of the algorithm model can be effectively improved through the improved model with the direction angle. The precision rate of the improved network model is 89.3%, which is much higher than the YOLO V3 algorithm, and also better than the YOLO V3 algorithm without direction detection. In terms of recall rate, the recall rate of the improved algorithm model reached 83.2%, which is still the best result among all algorithms. The F1-score of its algorithm also increased from 84.7% to 86.1%. The model designed in the article improves the detection accuracy of the direction angle to the target, but due to the increased complexity of the set model, the algorithm reasoning speed is reduced. However, the overall difference is not big, and it can meet the target detection task of UAV aerial photography vehicles.

On the whole, the improved algorithm model in the article can effectively improve the detection accuracy of the algorithm model without greatly sacrificing the detection rate, and can also improve the effect of small target detection of the model, which proves the effectiveness of the improved model.