IMPLEMENTATION OF CONVOLUTIONAL NEURAL NETWORK FOR PEDESTRIAN DETECTION IN AUTONOMOUS VEHICLE SYSTEM

• Main Author: Naufhal Dhiaegana, Renjira
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49926
• Institution: Institut Teknologi Bandung

Pedestrian detection is one of the many tasks that must be completed by the autonomous vehicle system. Nowadays, the use of convolutional neural networks (CNN) greatly affects object detection performance. You Only Look Once object detector (YOLO) is one of the object detector that can detect objects in real-time. The tradeoffs between accuracy and speed is a major problem in real-time object detection because the reduction of input resolution to the model eliminates object features. To overcome this problem, re-detection of areas containing most pedestrian objects, especially in the center of the image, is carried out. The pieces of the image are re-detected and the results of the detection are combined with the full image detection. Duplication of detection results is resolved by performing NMS (Non Maximum Suppression). The models used are YOLO version four as well as the tiny version, trained against the CrowdHuman dataset which can produce state-of-the-art detection performance (high real-time speed and mAP) for human object detection cases. Model performance is tested with the CrowdHuman dataset to find the best model from the training by looking at mAP (mean average precision). The best model is used in testing the object detection program with double detection. The state-of-the-art models YOLOv4 and YOLOv4-tiny as real-time models with high mAP were also tested. Test data from YouTube is used to determine the model's performance on cases in Indonesia. The use of a single model is compared with multiple models combined to determine the effect of double detection. From the test results, the double detection gave a good enough impact for detection in the tiny model. However, there are no significant results for detection in the YOLOv4 model. For the most optimal model, the YOLOv4-tiny-416-double model achieved 69.02 mAP compared to the one-time detection model YOLOv4-320 which reached 68.86 mAP and with a speed of 13.13 percent faster at 41 FPS.