graduation

SEU graduation project

Document description:

Main code file:

ui-t.py is the final running code

ui-yolov5 design source code for running interface

lane.py is the lane line detection code

Minor code

(1) Kitti dataset preprocessing code

d1modifyann.py

d2_txt2xml.py

d3_xml2txt.py

d4_divdataset.py

(2) yolov5 training, testing, etc.

detect.py detects code for yolov5

train.py is the training code

val.py is the model verification code

Other

The data folder is used to save data, and the yaml files are used to process the dataset.

The model folder is used to save the model, and the yaml files are used to save the model's structure, and so on.

The run folder is used to save the results of training, validation, and testing.

The icon folder is used to save the icons that the program needs to use.

The other folders are the folders that come with the Yolov5 model.

Lane Detection:

As shown in Figure 3-1 detection algorithm diagram, the input image will output the lane line information and detection result image after 3 steps. image.

The first step is preprocessing, after the input image is filtered by color filtering, L-channel filtering and Sobel's algorithm with a total of 3 thresholds, a grayscale binary image will be obtained. After the input image is filtered by color filtering, L-channel filtering and Sobel algorithm with three thresholds, a gray-scale binary image is obtained; the gray-scale binary image is transformed by IPM inverse perspective, and the view angle is transformed from front view to "bird's eye view"; the gray-scale binary image is transformed by IPM inverse perspective, and the view angle is transformed by IPM inverse perspective. The gray-scale binary image will be obtained after IPM inverse perspective transformation, and the viewpoint will be changed from front view to "bird's-eye view"; the bird's-eye view will be processed by the region of interest to obtain the cleaner lane line binary image.

The sliding window stage utilizes the pre-processed binary map to obtain the lane line information. Based on the previous binary map, first obtain The initial histogram is obtained to determine the initial position of the sliding window; the sliding window searches for the left and right lane lines according to the initial position. The sliding window is then slid according to the initial position to search for the left and right lane lines respectively; after that, the valid pixel points are obtained to fit the left and right lane lines.

In the post-processing part, the radius of curvature of the lane lines is calculated by using the coefficients of the fitted curves in combination with the multi-frame averaging method, The post-processing part utilizes the fitted curve coefficients and the multi-frame averaging method to calculate the radius of curvature of the lane lines, the sparsity, etc., and maps the fitted curves and lane line areas to the original image.

Advanced Sliding-Window

The basic sliding window makes it difficult to effectively detect lane lines in some situations. If the current lane is a sharp curve, the sliding window will use the previous window position to cause it to deviate from the lane line. Therefore, a more efficient method of finding the lane line is used.

The following figure is the result of the 2 different sliding-window methods.

https://github.com/z-z-n/The-Vision-based-Driver-Assistance-System/assets/65029895/448ad4dd-a159-4fdf-a086-face54a5ec43

Vehicle (Target) Detection:

In terms of target detection, I propose two CA hybrid attention improvement methods for the YOLOV5s model: YOLOv5s-CA and YOLOv5s-C3CA. YOLOv5s-CA and YOLOv5s-C3CA, respectively, to enhance the feature extraction ability of the model for the key regions of the image.

Experiments on the KITTI autonomous driving image dataset show that the introduction of the CA attention mechanism can effectively improve the performance of the models, and YOLOv5s-C3CA outperforms the basic YOLOv5s and YOLOv5s-CA models in terms of overall performance: without significantly increasing the complexity of the model and the detection time, it improves 1% and 2% of the mAP and recall R, respectively, compared to the YOLOv5s model. and 2% improvement in mAP and recall R respectively compared to the YOLOv5s model without significant increase in model complexity and time, especially in the detection of cyclists and pedestrians, the YOLOv5s-CA model also has some performance improvement compared to the YOLOv5s model and achieves the best results in the detection of vehicles, both in terms of the detection ability of the small targets and the confidence level.

Lane Detection + Vehicle Detection:

Combining lane line detection and vehicle detection with the YOLOv5-CA model.

https://github.com/z-z-n/The-Vision-based-Driver-Assistance-System/assets/65029895/1c6fc467-8ba5-4e4f-8116-3971a515565b