Adaptive Sliding-Window Strategy for Vehicle Detection in Highway Environments

Vehicle Detection Using Local Size-Specific Classifiers

“…To accurately detect vehicles, we should effectively address three major challenges: location ambiguity, size ambiguity, and appearance variation of vehicles [1], [2]. These challenges have been overcome by sliding window based detection methods [2]- [5] because of their superior performance [6], [7] compared with other approaches employing branch-andbound search [8], [9] and coarse-to-fine search [10]- [12]. Recently, more sophisticated techniques have also been proposed based on convolutional neural networks [13], [14] and 3D model structures [12], [15], [16].…”

“…Although considerable progress has been made [2], [5], sliding window schemes still encounter many problems when dealing with appearance variation of vehicles. To handle severe appearance variation, conventional approaches use an image classifier learned from a large-scale dataset.…”

“…2). In existing methods [2]- [6], normalization sizes are set manually based on expert's experiences, and therefore detection accuracy varies unstably according to specific scene conditions.…”

“…For a given scene, we first generate semantic region models (SRMs) in Fig. 1 (a) by analyzing the movement patterns of vehicles [2] and then register a set of size models for every SRM. Next, using training samples whose sizes are normalized based on the obtained size models, we construct more tailored image-classifiers, namely, local size-specific classifiers (LSCs).…”

“…In addition, LSCs help systems to operate detectors with lower computational costs. In conventional techniques [2]- [6], because appearance and size models are built independently, redundant combinations between them (e.g., a sedan appearance and a bus size) are also exhaustively analyzed. In contrast, our method avoids such unnecessary computations by jointly considering appearance and size information using LSCs.…”