An active learning system which generates training sequences comparable to randomly labeled video sequences has potential use in vehicle tracking

22 Oct 2024

UP Diliman

The development of an efficient and reliable transport system has a direct impact on the economy of the country. One of the ways that we can improve traffic management and monitoring is to make use of CCTVs, analyze the traffic video data and provide useful traffic information for better monitoring and management. This paves the way for what we call Intelligent Transportation Systems (ITS) where sophisticated algorithms such as machine learning and AI agents provides services to model and manage traffic systems, make them safer, more coordinated and makes efficient use of the transportation network. To enable the use of machine learning systems for ITS, there is a need to train the system on huge amounts of traffic data. The training data are labeled video data, the labels contain the necessary information that the machines need to learn, and it traditionally done manually consuming a lot of time and effort to develop.

The active learning framework is based on the idea that an algorithm can be used to determine which data is more useful and more informative to the machine learning system, so that only those need to be labeled manually. The idea is that the machine learning system will identify the video sequences that have low confidence scores, the active learning model will select the most informative of those sequences with low confidence and they will be manually labeled. After labeling those sequences will be included in the training data of the machine learning, improving the performance and also improving the identification of low
confidence sequences and the process repeats itself. After several iterations machine learning system is getting better at automatically labeling the video sequences and the active learning agent facilitates this improvement. The ultimate goal is to have a machine learn to automatically label unlabeled video sequences thus improving the training process as new data is generated.

Based on our small experiment, using an active learning system generates training sequences that are comparable with systems that learn from randomly labeled video sequences. In terms of traffic information such as vehicle tracking systems, the machine learning model trained using active learning frameworks performs slightly better than systems that rely on the same amount of randomly selected training data.

Authors: Adonais Ray Maclang, Miguel Lorenzo Orante, Rennuel Don Salvador, Dale Joshua Del Carmen, Rhandley D. Cajote (Electrical and Electronics Engineering Institute, University of the Philippines Diliman)

Read the paper: https://ieeexplore.ieee.org/document/10322372

An active learning system which generates training sequences comparable to randomly labeled video sequences has potential use in vehicle tracking

The development of an efficient and reliable transport system has a direct impact on the economy of the country. One of the ways that we can improve traffic management and monitoring is to make use of CCTVs, analyze the traffic video data and provide useful traffic information for better monitoring and management. This paves the way for what we call Intelligent Transportation Systems (ITS) where sophisticated algorithms such as machine learning and AI agents provides services to model and manage traffic systems, make them safer, more coordinated and makes efficient use of the transportation network. To enable the use of machine learning systems for ITS, there is a need to train the system on huge amounts of traffic data. The training data are labeled video data, the labels contain the necessary information that the machines need to learn, and it traditionally done manually consuming a lot of time and effort to develop.

The active learning framework is based on the idea that an algorithm can be used to determine which data is more useful and more informative to the machine learning system, so that only those need to be labeled manually. The idea is that the machine learning system will identify the video sequences that have low confidence scores, the active learning model will select the most informative of those sequences with low confidence and they will be manually labeled. After labeling those sequences will be included in the training data of the machine learning, improving the performance and also improving the identification of low
confidence sequences and the process repeats itself. After several iterations machine learning system is getting better at automatically labeling the video sequences and the active learning agent facilitates this improvement. The ultimate goal is to have a machine learn to automatically label unlabeled video sequences thus improving the training process as new data is generated.

Based on our small experiment, using an active learning system generates training sequences that are comparable with systems that learn from randomly labeled video sequences. In terms of traffic information such as vehicle tracking systems, the machine learning model trained using active learning frameworks performs slightly better than systems that rely on the same amount of randomly selected training data.

Authors: Adonais Ray Maclang, Miguel Lorenzo Orante, Rennuel Don Salvador, Dale Joshua Del Carmen, Rhandley D. Cajote (Electrical and Electronics Engineering Institute, University of the Philippines Diliman)

Read the paper: https://ieeexplore.ieee.org/document/10322372