Learning-based scene recognition with monocular camera for light-rail system

This paper is on scene recognition for a light railway vehicle safety system using a new patch-based approach for key frame identification. The approach is different from those conventional approaches using for example SIFT, SURF, BRIEF, or ORB for individual frame recognition. We propose a new unsupervised and learning-based key region detection method. The proposed method contains two parts. In the offline part, the key regions with discriminative information are identified from single reference sequence captured by monocular camera with unsupervised method. The discrimination power for a region is defined as the difference between this region and all other regions in the sequence. Regions having significant outstanding appearance are regarded as key regions. Binarization and greedy algorithm are used to choose key regions and discriminative patterns with low correlation. The key frames are key checking positions of the video path, whilst all other frames are tracked by matching approaches with substantially reduced computation. In the online part, each live frame is used initially to find the most nearby key frame, and the computation power of the subsequent detection is substantially reduced by looking for the next key frame with the frame by frame tracking procedure. Practical field tests were done on real data of the light railway system in Hong Kong. Results of these experimental tests show that the approach can identify almost 100% pre-recorded scene along railway paths with pedestrians. The approach has shown better performance over conventional approaches using some standard video sequences for scene recognition.