I have implemented the YOLOv4 algorithm to create custom trained weights for desired object detections(in our case Classes detected are: Car, Vehicle, Stop Signs, Person, Animals). and sliding window technique is merged to detect lanes.
Steps Involved in Implementing Custom trained YOLOv4 Detection in Google CoLab!
- Build Darknet
- Perform Detections with Darknet and YOLOv4 on Pre-trained weights
- Training a Custom YOLOv4 Object Detector in the Cloud
- Gather and Label Custom Dataset
- Train Custom Object Detector
Steps Involved in Lane detection!
- Pre-processing of Image
- Perspective transform
- Applying Sliding Windows Technique.
- Curve fitting
- Inverse perspective transform
- Plotting curve on input image/frame.
Input Image
Object Detection through Custom trained YOLOv4
Testing our object detection model on IISc campus:
Input:
Output:
Link to full Video: https://user-images.githubusercontent.com/111170719/206981021-ce4624af-9dc6-489a-b2ee-f48bda1bfbbc.mp4
Testing our object detection model on Random Delhi Highway YouTube video:
Input:
Output:
Link to full Video: https://user-images.githubusercontent.com/111170719/207027449-07eaaf4d-de20-4d43-bb5b-9dd183e48bdc.mp4
Testing our Lane detection and object detection model on Random sample video from YouTube:
Link to full Video: https://user-images.githubusercontent.com/111170719/206981713-f24840b8-fff4-4e2c-a71f-89946065f8de.mp4
Quantitative Evaluation:
Plot of Loss Curve with Iterations-
NOTE: Due to limitation on GPU usage for training on Google colab the loss curve is discontinous in between
FUTURE PLANS: THE MODEL WILL BE TRAINED ON MOTORCYCLE AND TRAFFIC SIGNAL DATASETS ALSO.